{"total":2584,"limit":100,"offset":0,"statements":[{"measure":"LM-2025-0002","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"The model proactively refuses high-risk tasks.","quote":"to address these risks, we have implemented a multi-layered approach to safety, including proactive refusals of high-risk tasks, confirmation prompts before critical actions, and active monitoring systems to detect and mitigate potential threats.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0003","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"The model asks for confirmation before critical actions.","quote":"to address these risks, we have implemented a multi-layered approach to safety, including proactive refusals of high-risk tasks, confirmation prompts before critical actions, and active monitoring systems to detect and mitigate potential threats.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0004","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Active monitoring systems detect and mitigate potential threats.","quote":"to address these risks, we have implemented a multi-layered approach to safety, including proactive refusals of high-risk tasks, confirmation prompts before critical actions, and active monitoring systems to detect and mitigate potential threats.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0005","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"The model adheres to a policy of human-in-the-loop safeguards across tasks and actions.","quote":"we aim to mitigate potential risks to users and others by encouraging the model to adhere to this policy of human-in-the-loop safeguards across tasks and actions (detailed in the risk mitigation section below).","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0006","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Operator users are bound by OpenAI usage policies, which apply universally to OpenAI services.","quote":"operator users are bound by openai usage policies 4 , which apply universally to openai services and are designed to ensure safe and responsible usage of ai technology.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0007","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Guidelines clarify how usage policies apply to Operator, emphasizing prohibited uses.","quote":"as part of this release, we are publishing guidelines to clarify how those usage policies apply to operator, explicitly emphasizing that operator should not be used to: facilitate or engage in illicit activity, including compromising the privacy of others, exploiting and harming children, or developing or distributing illicit substances, goods, or services, defraud, scam, spam or intentionally deceive or mislead others, including using operator to impersonate others without consent or legal right, falsely representing to others the extent to which they are engaging with an agent, or creating or employing deception or manipulation to inflict financial loss on others, engage in regulated activity without complying with applicable laws and regulations, including the use of operator to automat","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0008","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Operator is trained to refuse certain harmful tasks.","quote":"at the model level, operator is trained to refuse certain harmful tasks.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0009","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Operator is restricted from navigating to websites that could enable potentially harmful or illicit activities prohibited by OpenAI's usage policies.","quote":"at the system level, we restrict operator from navigating to websites that could enable potentially harmful or illicit activities that are prohibited by openai’s usage policies.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0010","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Automated and human review monitors for potential abuse and takes appropriate action with users who violate policies.","quote":"at the post-deployment stage, we will leverage automated and human review to monitor for potential abuse and take appropriate action with users who violate our policies.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0011","kind":"applied","model_version":"MV-2025-0040","answers":13,"by":"provider","text":"The model asks the user for confirmations before finalizing actions that affect the state of the world.","quote":"to further reduce the chance of model mistakes causing harm, we aim to have the model ask the user for confirmations before finalizing actions that affect the state of the world (e.g., before completing a purchase or sending an email).","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0012","kind":"applied","model_version":"MV-2025-0040","answers":14,"by":"provider","text":"The model proactively refuses certain high-risk tasks like banking transactions and making high-stakes decisions.","quote":"in this early stage, we decided to proactively refuse certain high-risk tasks like banking transactions and making high-stakes decisions, as per the risky task policy described in the policy cr","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0013","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"On certain websites, operator automatically pauses execution when the user becomes inactive or navigates away from the page.","quote":"in these scenarios, we require that the user supervise operator’s actions by automatically pausing execution when the user becomes inactive or navigates away from the page.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0014","kind":"applied","model_version":"MV-2025-0040","answers":16,"by":"provider","text":"A prompt injection monitor supervises execution of operator and pauses execution if a suspected prompt injection is detected on the screen.","quote":"on top of the model mitigations, we added a prompt injection monitor that is able to supervise execution of operator and will pause execution if a suspected prompt injection is detected on the screen (see figure 2).","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0015","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Operator safety checks are carried over into the API, providing increased visibility and early warning for potentially malicious instructions.","quote":"prompt injection and sensitive domain safety checks : we’ve carried over these operator safety checks into the api, providing increased visibility and early warning for potentially malicious instructions.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0016","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"A Docker application is provided for safer implementation, encouraging developers to use isolated environments.","quote":"containerized starter setup : we provide an easy-to-use docker application for safer implementation, encouraging developers to use isolated environments.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0017","kind":"applied","model_version":"MV-2025-0040","answers":null,"by":"provider","text":"Detection of potential policy violations, including jailbreaks, larger-scale misuse or suspicious patterns of API usage, has been expanded.","quote":"enhanced monitoring and enforcement : we’ve expanded detection of potential policy violations, including jailbreaks, larger-scale misuse or suspicious patterns of api usage.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-01-23","grade":"A","verified":true,"source":{"url":"https://openai.com/index/operator-system-card/","publisher":"OpenAI","date":"2025-01-23","type":"lab_disclosure","title":"Operator system card","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0040"},{"measure":"LM-2025-0137","kind":"applied","model_version":"MV-2025-0037","answers":null,"by":"provider","text":"At the critical risk threshold, access is strictly limited to a small number of experts.","quote":"Access is strictly limited to a small number of experts, alongside security protections to prevent hacking or exfiltration insofar as is technically feasible and commercially practicable.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://web.archive.org/web/20260303173438/https://ai.meta.com/static-resource/meta-frontier-ai-framework/","publisher":"Meta","date":"2025-02-01","type":"lab_disclosure","title":"Frontier AI Framework v1.0","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0037"},{"measure":"LM-2025-0138","kind":"applied","model_version":"MV-2025-0037","answers":null,"by":"provider","text":"At the critical risk threshold, security protections are applied to prevent hacking or exfiltration insofar as is technically feasible and commercially practicable.","quote":"Access is strictly limited to a small number of experts, alongside security protections to prevent hacking or exfiltration insofar as is technically feasible and commercially practicable.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://web.archive.org/web/20260303173438/https://ai.meta.com/static-resource/meta-frontier-ai-framework/","publisher":"Meta","date":"2025-02-01","type":"lab_disclosure","title":"Frontier AI Framework v1.0","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0037"},{"measure":"LM-2025-0139","kind":"applied","model_version":"MV-2025-0037","answers":null,"by":"provider","text":"For models at the high risk threshold, access is limited to a core research team, alongside security protections to prevent hacking or exfiltration.","quote":"7 Access is limited to a core research team, alongside security protections to prevent hacking or exfiltration.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://web.archive.org/web/20260303173438/https://ai.meta.com/static-resource/meta-frontier-ai-framework/","publisher":"Meta","date":"2025-02-01","type":"lab_disclosure","title":"Frontier AI Framework v1.0","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0037"},{"measure":"LM-2025-0047","kind":"applied","model_version":"MV-2025-0038","answers":null,"by":"provider","text":"A monitoring system to detect illicit use of instrumental reasoning capabilities in a model's chain-of-thought and use it in high-stakes deployments where there is significant risk of the model undermining human control.","quote":"Automated monitoring: a monitoring system to detect illicit use of instrumental reasoning capabilities in a model’s chain-of-thought and use it in high-stakes deployments where there is significant risk of the model undermining human control.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/updating-the-frontier-safety-framework/Frontier%20Safety%20Framework%202.0.pdf","publisher":"Google DeepMind","date":"2025-02-01","type":"lab_disclosure","title":"Frontier Safety Framework v2.0","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0038"},{"measure":"LM-2025-0153","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Anthropic's safeguards team runs continuous classifiers to monitor prompts and outputs for harmful use cases that violate the acceptable use policy.","quote":"Our Safeguards team also runs continuous classifiers to monitor prompts and outputs for harmful use cases that violate our AUP.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0142","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Anthropic is accelerating the development and deployment of targeted classifiers as part of enhanced ASL-2 safety measures.","quote":"In light of these findings, we are proactively enhancing our ASL-2 safety measures by accelerating the development and deployment of targeted classifiers and monitoring systems.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0140","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Anthropic is accelerating the development and deployment of monitoring systems as part of enhanced ASL-2 safety measures.","quote":"In light of these findings, we are proactively enhancing our ASL-2 safety measures by accelerating the development and deployment of targeted classifiers and monitoring systems.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0143","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Comprehensive monitoring systems and classifier-based interventions are implemented across key areas.","quote":"To address this, we’ve implemented comprehensive monitoring systems and classifier-based interventions across key areas, advancing responsible deployment while maintaining the model’s enhanced capabilities.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2026-0004","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Harmlessness training is a pre-deployment defense.","quote":"Pre-deployment defenses include harmlessness training and updating the computer use system prompt with language encouraging acceptable use.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0154","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"The computer use system prompt is updated with language encouraging acceptable use.","quote":"Pre-deployment defenses include harmlessness training and updating the computer use system prompt with language encouraging acceptable use.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0155","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Post-deployment defenses may include classifiers that flag potential harmful behavior by summarizing and classifying exchanges to identify misuse.","quote":"Post-deployment defenses may include leveraging 12 classifiers that flag potential harmful behavior by summarizing and classifying exchanges to identify misuse.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0156","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Enforcement action is taken against accounts found to be in violation of the usage policy, including user warnings, a system prompt suffix intervention, removing computer use capability, and banning an account or organization altogether.","quote":"We also take enforcement action against accounts found to be in violation of our Usage Policy, with actions including user warnings, a system prompt suffix intervention, removing computer use capability, and banning an account or organization altogether.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0157","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"A streaming completion classifier trained to detect and mitigate harmful content within chains of thought operates in real-time, analyzing content within thinking tags.","quote":"We implemented and deployed a streaming completion classifier trained to detect and mitigate harmful con- tent within chains of thought.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0146","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"For content identified by the streaming classifier as potentially harmful, encryption is employed as the intervention mechanism to prevent potentially harmful content from being exposed to the user while maintaining model functionality.","quote":"For content identified by the streaming classifier as potentially harmful, we employ encryption as the in- tervention mechanism, which is intended to prevent potentially harmful content from being exposed to the user while maintaining model functionality.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2026-0549","kind":"applied","model_version":"MV-2025-0039","answers":17,"by":"provider","text":"Classifiers are deployed to intervene and halt the model mid-execution if a prompt injection attack is detected.","quote":"We implemented a number of mitigations to reduce the risk of prompt injection attacks, including training the model to avoid this type of undesirable behavior during RL and deploying classifiers to intervene and halt the model mid-execution if a prompt injection attack is detected.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0149","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Automated classifiers detected the special-casing pattern in training transcripts as part of the broader safety monitoring framework.","quote":"However, our automated classifiers successfully detected this pattern in training transcripts as part of our broader safety monitoring framework.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0150","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Partial mitigations were implemented before launch to reduce the special-casing behavior.","quote":"After detection, we characterized the behavior and implemented partial mitigations before launch.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0151","kind":"applied","model_version":"MV-2025-0039","answers":null,"by":"provider","text":"Anthropic is proactively enhancing ASL-2 safety measures by accelerating the development and deployment of targeted classifiers and monitoring systems.","quote":"We are therefore proactively enhancing our ASL-2 safety measures by accelerating the development and deployment of targeted classifiers and monitoring systems.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/9ff93dfa8f445c932415d335c88852ef47f1201e.pdf","publisher":"Anthropic","date":"2025-02-01","type":"lab_disclosure","title":"Claude Sonnet 3.7","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0039"},{"measure":"LM-2025-0131","kind":"applied","model_version":"MV-2025-0034","answers":null,"by":"provider","text":"Documents published alongside each new model deployment, summarising the model’s intended use cases, performance on public benchmarks, and the responsible practices conducted to ensure safety.","quote":"Model Cards: Documents published alongside each new model deployment, summarising the model’s intended use cases, performance on public benchmarks, and the responsible practices conducted to ensure safety.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-06","grade":"A","verified":true,"source":{"url":"https://www.g42.ai/application/files/9517/3882/2182/G42_Frontier_Safety_Framework_Publication_Version.pdf","publisher":"G42","date":"2025-02-06","type":"lab_disclosure","title":"Frontier AI Safety Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0034"},{"measure":"LM-2025-0132","kind":"applied","model_version":"MV-2025-0034","answers":null,"by":"provider","text":"Designated pathways for users to report instances of concerning or harmful behavior in violation of company policy to relevant g42 personnel.","quote":"Incident Reporting Channels: Designated pathways for users to report instances of concerning or harmful behavior in violation of company policy to relevant G42 personnel.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-06","grade":"A","verified":true,"source":{"url":"https://www.g42.ai/application/files/9517/3882/2182/G42_Frontier_Safety_Framework_Publication_Version.pdf","publisher":"G42","date":"2025-02-06","type":"lab_disclosure","title":"Frontier AI Safety Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0034"},{"measure":"LM-2025-0133","kind":"applied","model_version":"MV-2025-0034","answers":null,"by":"provider","text":"Training programs for new and existing personnel on best practices in information security consistent with the measures described in the security mitigation levels.","quote":"Information Security Training: Training programs for new and existing personnel on best practices in information security consistent with the measures described in the Security Mitigation Levels.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-06","grade":"A","verified":true,"source":{"url":"https://www.g42.ai/application/files/9517/3882/2182/G42_Frontier_Safety_Framework_Publication_Version.pdf","publisher":"G42","date":"2025-02-06","type":"lab_disclosure","title":"Frontier AI Safety Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0034"},{"measure":"LM-2025-0134","kind":"applied","model_version":"MV-2025-0034","answers":null,"by":"provider","text":"Regular tests of ai models for robustness against attempts to manipulate or corrupt their output.","quote":"Model Robustness Testing: Regular tests of AI models for robustness against attempts to manipulate or corrupt their output.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-06","grade":"A","verified":true,"source":{"url":"https://www.g42.ai/application/files/9517/3882/2182/G42_Frontier_Safety_Framework_Publication_Version.pdf","publisher":"G42","date":"2025-02-06","type":"lab_disclosure","title":"Frontier AI Safety Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0034"},{"measure":"LM-2025-0135","kind":"applied","model_version":"MV-2025-0034","answers":null,"by":"provider","text":"Off-cycle review catches anomalies missed in real-time, assessing all stored interactions for unusual behaviors.","quote":"Asynchronous Monitoring: This off - cycle review catches anomalies missed in real-time, assessing all stored interactions for unusual behaviors.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-06","grade":"A","verified":true,"source":{"url":"https://www.g42.ai/application/files/9517/3882/2182/G42_Frontier_Safety_Framework_Publication_Version.pdf","publisher":"G42","date":"2025-02-06","type":"lab_disclosure","title":"Frontier AI Safety Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0034"},{"measure":"LM-2025-0136","kind":"applied","model_version":"MV-2025-0034","answers":null,"by":"provider","text":"For new frontier-level models, implement phased rollouts, starting with limited access to trusted users, with full deployment only after exhaustive risk assessments.","quote":"Controlled Rollout: For new frontier- level models, implement phased rollouts, starting with limited access to trusted users, with full deployment only after exhaustive risk assessments.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-06","grade":"A","verified":true,"source":{"url":"https://www.g42.ai/application/files/9517/3882/2182/G42_Frontier_Safety_Framework_Publication_Version.pdf","publisher":"G42","date":"2025-02-06","type":"lab_disclosure","title":"Frontier AI Safety Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0034"},{"measure":"LM-2025-0124","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"Guardrails, called safety modes, let customers control model outputs, such as selecting \"contextual\" or \"strict\" safety mode.","quote":"Guardrails, or what we call Safety Modes , are a feature that allows Cohere customers to exercise more precise control over model outputs.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0125","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"Where Cohere has direct visibility into model use during deployment, it monitors for malicious attempts to prompt models for harmful outputs and revokes access from accounts that abuse its systems.","quote":"Where Cohere has direct visibility into the use of its models during deployment, we use that visibility to monitor for malicious attempts to prompt our models for harmful outputs, revoking access from accounts that abuse our systems.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0126","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"Blocklists, custom classifiers, prompt injection guard filters, and human review are used to detect and intercept attempts to create unsafe outputs.","quote":"● Blocklists, custom classifiers, and prompt injection guard filters, and human review to detect and intercept attempts to create unsafe outputs","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0127","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"A SIEM system leveraging heuristics and advanced detection capabilities identifies potential threats.","quote":"● Security Information and Event Management (SIEM) system leveraging heuristics and advanced detection capabilities to identify potential threats","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0128","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"Air-gapped safeguards prevent lateral movement and unintended network calls across environments, and kernel-based LLMs prevent the leaking of shared memories or buffers that could expose sensitive data.","quote":"● “Air-gapped\" safeguards to prevent lateral movement and unintended network calls across environments and kernel-based LLMs to prevent the leaking of 3) Risk Assurance Mechanisms Assurance mechanisms are measures that provide confidence that risk mitigations are effective and expected risk mitigation objectives have been met.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0129","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"Safety classifiers and human review detect and intercept attempts to create unsafe outputs.","quote":"● Safety classifiers and human review to detect and intercept attempts to create unsafe outputs","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0130","kind":"applied","model_version":"MV-2025-0033","answers":null,"by":"provider","text":"Human-interpretable explanations of outputs are provided.","quote":"● Human-interpretable explanation of outputs","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-07","grade":"A","verified":true,"source":{"url":"https://cohere.com/security/the-cohere-secure-ai-frontier-model-framework-february-2025.pdf","publisher":"Cohere","date":"2025-02-07","type":"lab_disclosure","title":"Secure AI Frontier Model Framework","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0033"},{"measure":"LM-2025-0107","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Guardrails applied to the input to the LLM, which can reject, alter or mask potentially sensitive data before prompting the model.","quote":"d to prompt the LLM, or alter the relevant chunks (e.g., to mask potentially sensitive data).","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0108","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Guardrails applied to the input/output of custom actions (tools) that need to be called by the LLM.","quote":"• Execution rails are guardrails applied to input/output of the custom actions (a.k.a. tools), that need to be called by the LLM.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0109","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Guardrails applied to the output generated by the LLM that can reject the output, preventing it from being returned to the user, or alter it, e.g. removing sensitive data.","quote":"• Output rails are guardrails applied to the output generated by the LLM; an output rail can reject the output, preventing it from being returned to the user, or alter it (e.g., removing sensitive data).","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0110","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Blocks harmful visual outputs using a video content safety classifier and a face blur filter.","quote":"Cosmos post-Guard stage blocks harmful visual outputs using a video content safety classifier and a face blur filter.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0111","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Identifies near-duplicate content in training data and preserves the highest quality version for minimal data loss.","quote":"The risk of memorization is higher where data appears more than once in the training dataset.22 , so a deduplication step is normally used to identify near-duplicate content and preserves the highest quality version for minimal data loss.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0112","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Detects AI generated content in training datasets.","quote":"In addition to conventional approaches for measuring data quality we have developed and shared novel methods for detecting AI generated content in training datasets through HIVE23 as a NIM.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0113","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Filters training data to exclude examples that could result in capabilities that increase the likelihood of misuse, such as known CSAM images.","quote":"We also filter training data to exclude examples that could result in capabilities that increase the likelihood of misuse, such known CSAM images.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0114","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Subtle statistical or cryptographic markers embedded into the distributional properties of generated outputs to increase detectability of content generated by Nvidia models.","quote":"There are several opensource implicit watermarking tools that may be used to increase the detectability of content generated by NVIDIA models.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0115","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Visible or otherwise directly perceivable indicators such as logos, text overlays or perceptible patterns that denote AI-generated content to end-users; use is specific to downstream developers and not integrated into Nvidia's foundation models.","quote":"The use of explicit watermarks are specific to downstream developers.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0116","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Restricting the scope and use of a model by imposing capability or feature restrictions, such as limiting the types of inputs a model can process.","quote":"This is achieved by imposing capability or feature restrictions, such as limiting the types of inputs a model can process.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0117","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Models may be explicitly barred from prohibited applications through legal mechanisms such as Nvidia's end user license agreements for foundation models.","quote":"Additionally, models may be explicitly barred from prohibited applications through legal mechanisms such as NVIDIA’s End User License Agreements for foundation models.26","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0118","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Restricting advanced autonomy functions like self-assigning new sub-goals or executing long-horizon tasks, as well as tool-use functionalities like function calls and web browsing.","quote":"Another important strategy involves restricting advanced autonomy functions like self-assigning new sub-goals or executing long-horizon tasks, as well as tool-use functionalities like function calls and web browsing.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0119","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Ensuring only authorized users access the model through secure API keys and authentication protocols, performing know-your-customer (KYC) screenings for users with high output needs, and limiting access frequency by capping requests or instituting time-based quotas.","quote":"These include ensuring only authorized users access the model through secure API keys and authentication protocols, performing Know-Your-Customer (KYC) screenings for users with high output needs, and limiting access frequency by capping requests30 or instituting time-based quotas.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0120","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Detecting and blocking misuse attempts using algorithmic classifiers, which can limit unsafe queries, modify responses, or block users attempting to bypass safeguards.","quote":"This includes detecting and blocking misuse attempts using algorithmic classifiers, which can limit unsafe queries, modify responses, or block users attempting to bypass safeguards.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0121","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"When a model shows capabilities of frontier AI models pre deployment, access to model weights is initially restricted to essential personnel with rigorous security protocols in place.","quote":"When a model shows capabilities of frontier AI models pre deployment we will initially restrict access to model weights to essential personnel and ensure rigorous security protocols are in place.34","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0122","kind":"applied","model_version":"MV-2025-0031","answers":null,"by":"provider","text":"Measures to restrict at-will fine tuning of frontier AI models without safeguards in NeMo Customizer, reducing the options to retrain a model on data related to dangerous tasks or to reduce how often the model refuses potentially dangerous requests.","quote":"Measures will also be in place to restrict at-will fine tuning of frontier AI models without safeguards in NeMo customizer, reducing the options to retrain a model on data related to dangerous tasks or to reduce how often the model refuses potentially dangerous requests.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0123","kind":"applied","model_version":"MV-2025-0031","answers":0,"by":"provider","text":"Garak is used as a highest-priority assessment of models before release.","quote":"NVIDIA takes advantage of this and uses Garak as a highest-priority assessment of models before release.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-02-17","grade":"A","verified":true,"source":{"url":"https://images.nvidia.com/content/pdf/NVIDIA-Frontier-AI-Risk-Assessment.pdf","publisher":"Nvidia","date":"2025-02-17","type":"lab_disclosure","title":"Frontier AI Risk Assessment","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0031"},{"measure":"LM-2025-0091","kind":"applied","model_version":"MV-2025-0030","answers":null,"by":"provider","text":"All of Anthropic's models must meet the ASL-2 deployment standard, which reduces the prevalence of misuse and includes publication of model cards, enforcement of usage policy, harmlessness training such as Constitutional AI, automated detection mechanisms, vulnerability reporting channels, and a bug bounty for universal jailbreaks.","quote":"At present, all of our models must meet the ASL-2 Deployment and Security Standards.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-03-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/17310f6d70ae5627f55313ed067afc1a762a4068.pdf","publisher":"Anthropic","date":"2025-03-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.1","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0030"},{"measure":"LM-2025-0092","kind":"applied","model_version":"MV-2025-0030","answers":null,"by":"provider","text":"All of Anthropic's models must meet the ASL-2 security standard, which requires a security system that can likely thwart most opportunistic attackers and includes vendor and supplier security reviews, physical security measures, and secure-by-design principles.","quote":"The ASL-2 Security Standard requires a security system that can likely thwart most opportunistic attackers and includes vendor and supplier security reviews, physical security measures, and the use of secure-by-design principles.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-03-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/17310f6d70ae5627f55313ed067afc1a762a4068.pdf","publisher":"Anthropic","date":"2025-03-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.1","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0030"},{"measure":"LM-2025-0028","kind":"applied","model_version":"MV-2025-0028","answers":null,"by":"provider","text":"Conditional pre-training is applied as a mitigation.","quote":"conditional pre-training;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-10","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Lite-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-10","type":"lab_disclosure","title":"Gemini 2.0 Flash-Lite","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0028"},{"measure":"LM-2025-0029","kind":"applied","model_version":"MV-2025-0028","answers":null,"by":"provider","text":"Supervised fine-tuning is applied as a mitigation.","quote":"supervised fine-tuning;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-10","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Lite-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-10","type":"lab_disclosure","title":"Gemini 2.0 Flash-Lite","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0028"},{"measure":"LM-2025-0030","kind":"applied","model_version":"MV-2025-0028","answers":null,"by":"provider","text":"Reinforcement learning from human and critic feedback is applied as a mitigation.","quote":"reinforcement learning from human and critic feedback;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-10","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Lite-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-10","type":"lab_disclosure","title":"Gemini 2.0 Flash-Lite","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0028"},{"measure":"LM-2025-0031","kind":"applied","model_version":"MV-2025-0028","answers":null,"by":"provider","text":"Safety policies and desiderata are applied as a mitigation.","quote":"safety policies and desiderata;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-10","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Lite-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-10","type":"lab_disclosure","title":"Gemini 2.0 Flash-Lite","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0028"},{"measure":"LM-2025-0001","kind":"applied","model_version":"MV-2025-0028","answers":null,"by":"provider","text":"Product-level mitigations such as safety filtering are applied.","quote":"product-level mitigations such as safety filtering.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-10","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Lite-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-10","type":"lab_disclosure","title":"Gemini 2.0 Flash-Lite","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0028"},{"measure":"LM-2026-0238","kind":"applied","model_version":"MV-2025-0026","answers":0,"by":"provider","text":"Safeguards are built to prevent the models from assisting with high-risk queries related to building weapons.","quote":"We build safeguards to prevent our models from assisting with high-risk queries related to building weapons, and evaluate performance on those refusal policies as part of our safety process.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-15","grade":"A","verified":true,"source":{"url":"https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf","publisher":"OpenAI","date":"2025-04-15","type":"lab_disclosure","title":"Preparedness Framework, v2.0","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0026"},{"measure":"LM-2025-0028","kind":"applied","model_version":"MV-2025-0027","answers":null,"by":"provider","text":"Conditional pre-training is applied as a mitigation.","quote":"conditional pre-training;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-15","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-15","type":"lab_disclosure","title":"Gemini 2.0 Flash","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0027"},{"measure":"LM-2025-0029","kind":"applied","model_version":"MV-2025-0027","answers":null,"by":"provider","text":"Supervised fine-tuning is applied as a mitigation.","quote":"supervised fine-tuning;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-15","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-15","type":"lab_disclosure","title":"Gemini 2.0 Flash","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0027"},{"measure":"LM-2025-0030","kind":"applied","model_version":"MV-2025-0027","answers":null,"by":"provider","text":"Reinforcement learning from human and critic feedback is applied as a mitigation.","quote":"reinforcement learning from human and critic feedback;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-15","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-15","type":"lab_disclosure","title":"Gemini 2.0 Flash","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0027"},{"measure":"LM-2025-0031","kind":"applied","model_version":"MV-2025-0027","answers":null,"by":"provider","text":"Safety policies and desiderata are applied as a mitigation.","quote":"safety policies and desiderata;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-15","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-15","type":"lab_disclosure","title":"Gemini 2.0 Flash","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0027"},{"measure":"LM-2025-0001","kind":"applied","model_version":"MV-2025-0027","answers":null,"by":"provider","text":"Product-level mitigations such as safety filtering are applied.","quote":"product-level mitigations such as safety filtering.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-04-15","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-04-15","type":"lab_disclosure","title":"Gemini 2.0 Flash","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0027"},{"measure":"LM-2025-0091","kind":"applied","model_version":"MV-2025-0021","answers":null,"by":"provider","text":"All of Anthropic's models must meet the ASL-2 deployment standard, which includes publication of model cards, enforcement of usage policy, harmlessness training such as Constitutional AI, automated detection mechanisms, vulnerability reporting channels, and a bug bounty for universal jailbreaks.","quote":"At present, all of our models must meet the ASL-2 Deployment and Security Standards.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.2","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0021"},{"measure":"LM-2025-0092","kind":"applied","model_version":"MV-2025-0021","answers":null,"by":"provider","text":"All of Anthropic's models must meet the ASL-2 security standard, which requires a security system that can likely thwart most opportunistic attackers and includes vendor and supplier security reviews, physical security measures, and secure-by-design principles.","quote":"The ASL-2 Security Standard requires a security system that can likely thwart most opportunistic attackers and includes vendor and supplier security reviews, physical security measures, and the use of secure-by-design principles.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.2","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0021"},{"measure":"LM-2026-0004","kind":"applied","model_version":"MV-2025-0022","answers":10,"by":"provider","text":"Pre-deployment harmlessness training was implemented to address malicious computer use concerns.","quote":"to address these concerns, we implemented several safeguards, including pre-deployment measures such as harmlessness training and updating the computer use instructions to emphasize appropriate usage.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0093","kind":"applied","model_version":"MV-2025-0022","answers":10,"by":"provider","text":"Computer use instructions were updated to emphasize appropriate usage.","quote":"to address these concerns, we implemented several safeguards, including pre-deployment measures such as harmlessness training and updating the computer use instructions to emphasize appropriate usage.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0094","kind":"applied","model_version":"MV-2025-0022","answers":10,"by":"provider","text":"Additional monitoring of harmful behavior was implemented, with post-deployment action against violating accounts.","quote":"we implemented additional monitoring of harmful behavior and, post-deployment, can take action against accounts that violate our usage policy by adding system prompt interventions, removing computer capabilities, or completely banning accounts or organizations.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0083","kind":"applied","model_version":"MV-2025-0022","answers":11,"by":"provider","text":"Specialized reinforcement learning training helps the model recognize and avoid prompt injection manipulations.","quote":"we implemented several protective measures to combat prompt injection attacks, including specialized reinforcement learning training to help the model recognize and avoid these manipulations, and the deployment of detection systems that can halt the model’s execution when a potential injection attempt is identified.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0095","kind":"applied","model_version":"MV-2025-0022","answers":11,"by":"provider","text":"Detection systems can halt the model's execution when a potential injection attempt is identified.","quote":"we implemented several protective measures to combat prompt injection attacks, including specialized reinforcement learning training to help the model recognize and avoid these manipulations, and the deployment of detection systems that can halt the model’s execution when a potential injection attempt is identified.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0096","kind":"applied","model_version":"MV-2025-0022","answers":12,"by":"provider","text":"Post-deployment measures steer and detect malicious coding use.","quote":"we implemented several measures to combat malicious coding requests, including harmlessness training and post-deployment measures to steer and detect for malicious use.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0097","kind":"applied","model_version":"MV-2025-0022","answers":12,"by":"provider","text":"Additional safety monitoring ensures the measures are effective and can take additional action like banning accounts or organizations.","quote":"finally, we are deploying additional safety monitoring to ensure these measures are effective and can take additional action like banning accounts or organizations.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0098","kind":"applied","model_version":"MV-2025-0022","answers":null,"by":"provider","text":"Monitoring for problematic behaviors during training was increased and improved, including iterative classifier development and unsupervised investigative explorations.","quote":"we increased and improved our monitoring for problematic behaviors during training.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0099","kind":"applied","model_version":"MV-2025-0022","answers":null,"by":"provider","text":"Adjustments were made to reduce hacking vulnerabilities in training environments, environment instructions were modified to be more consistent with the reward signal, and the reward signal during reinforcement learning was adjusted to be more robust to reward hacking.","quote":"we made a number of adjustments to reduce hacking vulnerabilities in our training environments.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0022"},{"measure":"LM-2025-0100","kind":"applied","model_version":"MV-2025-0023","answers":null,"by":"provider","text":"ASL-3 safeguards are in place for Claude Opus 4, blocking all harmful responses when re-running biology-related prompts.","quote":"when re-running biology-related prompts with our new asl-3 safeguards in place, all harmful responses were blocked.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0023"},{"measure":"LM-2025-0104","kind":"applied","model_version":"MV-2025-0023","answers":29,"by":"provider","text":"A targeted dataset was introduced into Claude Opus 4's training in which the model would be prompted with prompts from the paper, but trained to give responses based on the behavior of Claude Sonnet 3.7, an earlier model not influenced by the alignment faking data.","quote":"to mitigate these hallucinations, we introduced a targeted dataset into claude opus 4’s training in which the model would be prompted with prompts from the paper, but trained to give responses based on the behavior of claude sonnet 3.7, an earlier model not influenced by the alignment faking data.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0023"},{"measure":"LM-2025-0105","kind":"applied","model_version":"MV-2025-0023","answers":null,"by":"provider","text":"Machine-readable canary strings are being added to publicly released transcripts to facilitate the exclusion of these data from future pretraining corpora.","quote":"going forward, we are also adding machine-readable canary strings to our publicly released transcripts to facilitate the exclusion of these data from future pretraining corpora.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0023"},{"measure":"LM-2025-0152","kind":"applied","model_version":"MV-2025-0023","answers":null,"by":"provider","text":"Users should exercise caution with instructions that invite high-agency behavior in contexts that could appear ethically questionable.","quote":"we recommend that users exercise caution with instructions like these that invite high-agency behavior in contexts that could appear ethically questionable.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-01","grade":"A","verified":true,"source":{"url":"https://www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64dd.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Claude Sonnet 4 and Opus 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0023"},{"measure":"LM-2025-0087","kind":"applied","model_version":"MV-2025-0020","answers":null,"by":"provider","text":"Unsafe, violent, or low-quality images were removed from the training data.","quote":"Safety and quality image filtering : removal of unsafe, violent, or low-quality images.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-20","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Imagen-4-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-05-20","type":"lab_disclosure","title":"Imagen 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0020"},{"measure":"LM-2025-0088","kind":"applied","model_version":"MV-2025-0020","answers":null,"by":"provider","text":"AI-generated images were removed from the training data to prevent the model from learning artifacts or biases.","quote":"Eliminating AI-generated images : removal of AI-generated images prevents the model from learning artifacts or biases that may be found in AI-generated images.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-20","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Imagen-4-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-05-20","type":"lab_disclosure","title":"Imagen 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0020"},{"measure":"LM-2025-0089","kind":"applied","model_version":"MV-2025-0020","answers":null,"by":"provider","text":"Deduplication pipelines were used and similar images were down-weighted to minimize the risk of outputs overfitting training data.","quote":"Deduplicating images : deduplication pipelines were utilized and similar images were down-weighted to minimize the risk of outputs overfitting training data.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-20","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Imagen-4-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-05-20","type":"lab_disclosure","title":"Imagen 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0020"},{"measure":"LM-2025-0090","kind":"applied","model_version":"MV-2025-0020","answers":null,"by":"provider","text":"Production filtering may be applied to minimize harmful outputs.","quote":"Post-training mitigations may include production filtering to minimize harmful outputs, and application of tools such as SynthID watermarking to reduce risks such as misinformation.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-20","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Imagen-4-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-05-20","type":"lab_disclosure","title":"Imagen 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0020"},{"measure":"LM-2026-0250","kind":"applied","model_version":"MV-2025-0020","answers":null,"by":"provider","text":"SynthID watermarking may be applied to reduce risks such as misinformation.","quote":"Post-training mitigations may include production filtering to minimize harmful outputs, and application of tools such as SynthID watermarking to reduce risks such as misinformation.","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-05-20","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Imagen-4-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-05-20","type":"lab_disclosure","title":"Imagen 4","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0020"},{"measure":"LM-2025-0027","kind":"applied","model_version":"MV-2025-0019","answers":null,"by":"provider","text":"Dataset filtering is applied as a mitigation.","quote":"Mitigations include, but are not limited to: ● dataset filtering;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-06-27","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Pro-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-06-27","type":"lab_disclosure","title":"Gemini 2.5 Pro","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0019"},{"measure":"LM-2025-0028","kind":"applied","model_version":"MV-2025-0019","answers":null,"by":"provider","text":"Conditional pre-training is applied as a mitigation.","quote":"● conditional pre-training;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-06-27","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Pro-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-06-27","type":"lab_disclosure","title":"Gemini 2.5 Pro","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0019"},{"measure":"LM-2025-0029","kind":"applied","model_version":"MV-2025-0019","answers":null,"by":"provider","text":"Supervised fine-tuning is applied as a mitigation.","quote":"● supervised fine-tuning;","source_index":0,"scope":null,"obligation":null,"clause":null,"date":"2025-06-27","grade":"A","verified":true,"source":{"url":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Pro-Model-Card.pdf","publisher":"Google DeepMind","date":"2025-06-27","type":"lab_disclosure","title":"Gemini 2.5 Pro","archived_url":null,"attachment":null},"model_version_id":"MV-2025-0019"}]}