AI Safety

What AI providers state their models reached in safety testing, in their own words, and what outside testers found.

Also: compare models side by side · statements that a capability level was crossed · providers · practices providers document · safety news.

What providers state

Each row is one statement in a provider's own safety document about a capability level of one of its models: the level as the provider names it, whether the provider says the model reached it, and the sentence that says so.

139 of 139

Model	Level, as the provider names it	The provider states	The provider's sentence
Muse Spark Meta	Moderate or lower Autonomy and AI research	Not reached	“We assess loss of control risk as “moderate or lower”.” source
Muse Spark Meta	High Chemical, biological, radiological, nuclear	Not determined	“Those evaluations identified elevated risks prior to mitigations, with Chemical and Biological capabilities assessed as likely reaching the “high risk” category under the Advanced AI Scaling Framework before safeguards were applied.” source
Muse Spark Meta	Moderate or lower Chemical, biological, radiological, nuclear	Not reached	“However, after applying mitigations, the residual risk of the model is that of “moderate or lower risk.”” source
Muse Spark Meta	Below the level required to meaningfully enable the cyber threat scenarios defin Cyber	Not reached	“Muse Spark’s offensive cyber capabilities remain below the level required to meaningfully enable the cyber threat scenarios defined in the Advanced AI Scaling Framework.” source
Muse Spark Meta	Moderate or lower Cyber	Not reached	“We assess cybersecurity risk as “moderate or lower” with respect to the threat scenarios outlined in Meta’s Advanced AI Scaling Framework.” source
Gemini 4 Argon Google DeepMind	TCL (Tracked Capability Level) Autonomy and AI research	Not determined	“In light of this, we assess that we cannot rule out the TCL.” source
Gemini 4 Argon Google DeepMind	TCL (Tracked Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the TCL for the chemical, radiological, and nuclear domains with reasonable confidence.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Deception and sabotage	Not reached	“we assess its overall capabilities to be below the CCL threshold.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Misuse and jailbreaks	Not reached	“Gemini 4 Argon demonstrates some ability to influence user beliefs and behaviors during one-on-one direct conversations in human behavioral studies, and may provide utility in the planning stage of a malicious influence operation but we assess its overall capabilities to be below the CCL threshold.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Other	Not reached	“Gemini 4 Argon's performance remains below our CCL threshold.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Autonomy and AI research	Not reached	“The model does not reach the CCL alert threshold.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the CCL for the CBRN domain with reasonable conﬁdence based on the results from our testing.” source
Gemini 3.7 Flash Google DeepMind	TCL (Tracked Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the TCL for the CBRN domain with reasonable conﬁdence based on the results from our testing.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Cyber	Not reached	“Gemini 3.7 Flash reaches the alert threshold for this CCL, but not the CCL.” source
Gemini 3.7 Flash Google DeepMind	TCL (Tracked Capability Level) Deception and sabotage	Not reached	“The model does not reach the TCL.” source
Gemini 3.5 Transcribe Live Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that none of Gemini 3.5 Translate, Gemini 3.5 Transcribe, or Gemini 3.5 Transcribe Live have meaningful new T/CCL-relevant capabilities or material increases in performance with respect to Frontier Safety compared to Gemini 3.1 Pro and Gemini 3.7 Flash, therefore we are conﬁdent that none of them are likely to reach any CCLs.” source
Gemini 3.5 Transcribe Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that none of Gemini 3.5 Translate, Gemini 3.5 Transcribe, or Gemini 3.5 Transcribe Live have meaningful new T/CCL-relevant capabilities or material increases in performance with respect to Frontier Safety compared to Gemini 3.1 Pro and Gemini 3.7 Flash, therefore we are conﬁdent that none of them are likely to reach any CCLs.” source
Gemini 3.5 Flash-Lite Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that, while Gemini 3.5 Flash-Lite excels at agents and coding, it does not have meaningful new capabilities or material increases in performance with respect to the domains outlined in our Frontier Safety Framework compared to Gemini 3.1 Pro; therefore, based on Gemini 3.1 Pro results, we are confident that Gemini 3.5 Flash-Lite is also unlikely to reach any CCLs.” source
Gemini 3.5 Audio Google DeepMind	Tracked or critical capability level (t/ccls) Other	Not reached	“they did not reach any Tracked or Critical Capability Levels (T/CCLs) outlined in our Frontier Safety Framework .” source
Gemini 3.1 Pro Google DeepMind	Uplift level 1 Chemical, biological, radiological, nuclear	Not reached	“We found that the model remains below alert thresholds for the CBRN, harmful manipulation, machine learning R&D, and misalignment CCLs.” source
Gemini 3.1 Pro Google DeepMind	Uplift level 1 Cyber	Not reached	“As previous models passed the alert threshold for cyber, we performed more additional testing in this domain on Gemini 3.1 Pro with and without Deep Think mode, and found that the model remains below the cyber CCL.” source
Gemini 3.1 Flash-Lite Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that Gemini 3.1 Flash-Lite is less capable than Gemini 3.1 Pro, therefore based on Gemini 3.1 Pro, we are confident that Gemini 3.1 Flash-Lite is also unlikely to reach any CCLs.” source
Gemini 3.1 Flash Audio Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that Gemini 3.1 Flash Audio is less capable than Gemini 3.1 Pro.” source
Gemini 2.5 Pro Google DeepMind	Machine Learning R&D Uplift Level 1 Autonomy and AI research	Not determined	“given Gemini 2.5 Pro Preview's poor median performance on RE-Bench relative to experts, the model likely lacks the necessary capabilities to automate or signiﬁcantly uplift any signiﬁcant fraction of the research process.” source
Gemini 2.5 Pro Google DeepMind	Instrumental Reasoning Level 1 Deception and sabotage	Not reached	“Instru- mental Reasoning Level 1 ✘ CCL not reached” source
Gemini 2.5 Pro Google DeepMind	Instrumental Reasoning Level 2 Deception and sabotage	Not reached	“Instru- mental Reasoning Level 2 ✘ CCL not reached” source
Gemini 2.5 Deep Think Google DeepMind	Uplift level 1 Chemical, biological, radiological, nuclear	Not determined	“Our assessment is that the model has enough technical knowledge in certain CBRN scenarios and stages to be considered at early warning.” source
Gemini 2.5 Deep Think Google DeepMind	Cyber Uplift Level 1 Cyber	Not determined	“CCL reached? No, but the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.” source
Gemini 2.5 Deep Think Google DeepMind	Uplift level 1 Cyber	Not determined	“For Cyber Uplift Level 1 , the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.” source
GPT-6.1 Terra OpenAI	High Autonomy and AI research	Not reached	“In AI Self-Improvement, GPT-6.1 Terra does not reach our High threshold.” source
GPT-6.1 Terra OpenAI	High Chemical, biological, radiological, nuclear	Reached	“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” source
GPT-6.1 Terra OpenAI	Critical Cyber	Reached	“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” source
GPT-5.5 Instant OpenAI	Below high capability Autonomy and AI research	Not reached	“GPT-5.5 Instant is considered below High Capability.” source
GPT-5.5 Instant OpenAI	High Autonomy and AI research	Not reached	“For AI Self-Improvement, evaluations of final checkpoints indicate that, like its predecessor models, GPT-5.5 Instant does not have a plausible chance of reaching a High threshold.” source
GPT-5.5 Instant OpenAI	High Chemical, biological, radiological, nuclear	Reached	“GPT-5.5 Instant is our first Instant model to be treated as High Capability in the Biological and Chemical domain, as well as in the Cybersecurity domain.” source
GPT-5.5 Instant OpenAI	High Cyber	Reached	“this is the first Instant model that we are treating as High capability in our Cybersecurity and Biological & Chemical Preparedness categories, and implementing appropriate safeguards.” source
GPT-5 OpenAI	High Chemical, biological, radiological, nuclear	Reached	“Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards.” source
GPT-5 OpenAI	High cyber risk Cyber	Not reached	“The gpt-5 model series does not meet the threshold for high cyber risk.” source
Claude Sonnet 4.5 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“All models tested remain below our pre-committed rule-out threshold of 0.6.” source
Claude Sonnet 4.5 Anthropic	ASL-3 Autonomy and AI research	Not determined	“Exhibited improved autonomy capabilities in software engineering and AI research tasks, though still below the ASL-4 rule-out threshold.” source

Previous · Page 3 of 4 · Next

What outside testers report

Each row is one result from a government evaluator, independent lab, academic or security firm that tested a model, quoted from where it is reported.

1940 of 1940

Model	Tester	Test	What was found
Claude Sonnet 4.6 Anthropic	Independent lab Independent lab · 2026-02-17	Vending-Bench 2 Deception and sabotage	“Andon reports that Sonnet 4.6 was comparably aggressive to Opus 4.6 in its business practices, including lying to suppliers and initiating price-ﬁxing in some cases, though it lacked Opus 4.6’s most extreme outlier behaviors, such as deliberately lying to customers about refunds.” source
Claude Sonnet 4.6 Anthropic	Gray Swan, UK AI Security Institute Independent lab · 2026-02-17	Agent Red Teaming (ART) benchmark Prompt injection and agents	“Gray Swan , an external research partner, evaluated our models using the Agent Red Teaming (ART) benchmark, 33 developed in collaboration with the UK AI Security Institute .” source
GPT-5 OpenAI	METR Independent lab · 2026-02-13	time horizon measurement using Triframe Autonomy and AI research	“Prior to my investigation, METR had already measured the time horizon of: - Opus 4.5 using ReAct - GPT-5 using Triframe4.” source
GPT-5 OpenAI	METR Independent lab · 2026-02-13	time horizon measurement using Codex Autonomy and AI research	“I additionally measured the time horizon of: - Opus 4.5 using Claude Code - GPT-5 using Codex.” source
GPT-5 OpenAI	METR Independent lab · 2026-02-13	cheating detection (LLM-based cheating scanner over all of the runs, with manual review of flagged runs) Other	“I also did cheating detection (running an LLM-based cheating scanner over all of the runs, and manually reviewing ones that the scanner flagged), and noted examples of the kinds of mistakes Claude Code and Codex scaffolds tend to make in long-horizon tasks.” source
Claude Opus 4.5 Anthropic	METR Independent lab · 2026-02-13	time horizon measurement using ReAct Autonomy and AI research	“Prior to my investigation, METR had already measured the time horizon of: - Opus 4.5 using ReAct” source
Claude Opus 4.5 Anthropic	METR Independent lab · 2026-02-13	time horizon measurement using Claude Code Autonomy and AI research	“I additionally measured the time horizon of: - Opus 4.5 using Claude Code” source
Claude Opus 4.5 Anthropic	METR Independent lab · 2026-02-13	time horizon measurement with a four times higher token budget of 32 million tokens Autonomy and AI research	“I decided to run another time horizon measurement with a four times higher token budget of 32 million tokens.” source
Claude Opus 4.5 Anthropic	METR Independent lab · 2026-02-13	cheating detection Other	“I also did cheating detection (running an LLM-based cheating scanner over all of the runs, and manually reviewing ones that the scanner flagged), and noted examples of the kinds of mistakes Claude Code and Codex scaffolds tend to make in long-horizon tasks.” source
Gemini 3 Pro Google DeepMind	FAR.AI Independent lab · 2026-02-10	attempt-to-persuade-eval (APE) Harmful content	“However, Gemini 3 Pro complies with almost all persuasion requests, including on mass murder, physical violence, torture, human trafficking, and child sexual abuse.” source
Gemini 3 Pro Google DeepMind	FAR.AI Independent lab · 2026-02-10	attempt-to-persuade eval (APE) Misuse and jailbreaks	“But Google’s Gemini 3 Pro complies with almost any persuasion request in our evaluation, without jailbreaking.” source
Gemini 3 Pro Google DeepMind	FAR.AI Independent lab · 2026-02-10	attempt-to-persuade eval (APE) with low reasoning setting Misuse and jailbreaks	“As shown above, the low reasoning setting leads to an attempt rate of ~90% on our "noncontroversially harmful" topic category, whereas high reasoning leads to a lower attempt rate, of ~25%.” source
Gemini 3 Pro Google DeepMind	FAR.AI Independent lab · 2026-02-10	manipulation propensity evaluation Deception and sabotage	“Our findings are consistent with Google’s Gemini 3 Pro Frontier Safety Framework , which reports increased manipulation propensity in Gemini 3 Pro relative to earlier models.” source
Gemini 2.5 Pro Google DeepMind	FAR.AI Independent lab · 2026-02-10	attempt-to-persuade-eval (APE) Misuse and jailbreaks	“Gemini 2.0 Flash, GPT-4o, Gemini 2.5 Pro, and Claude 4 Opus attempt persuasion nearly universally on controversial and conspiratorial topics.” source
Claude Opus 4.5 Anthropic	FAR.AI Independent lab · 2026-02-10	attempt-to-persuade-eval (APE) Misuse and jailbreaks	“On the positive side, GPT-5.1 and Claude Opus 4.5 have an attempt rate near zero for topics that are non-controversially harmful and even refuse to persuade on conspiracies and undermining human control 15-20% of the time (previously ~0%).” source
Claude Opus 4.5 Anthropic	FAR.AI Independent lab · 2026-02-10	attempt-to-persuade eval (APE) across reasoning settings Misuse and jailbreaks	“This pattern does not generalize: Both GPT-5.1 and Claude Opus 4.5 produces ~0% harmful attempts regardless of reasoning effort.” source
GPT-5.3-Codex OpenAI	Independent lab Independent lab · 2026-02-05	network attack simulation Cyber	“GPT- 5.3-Codex achieved (success rates compared to GPT-5.2-Codex, GPT-5.2-thinking in parentheses) average success rates of 86% (68%,100%) on Network Attack Simulation challenges, 72% (75%, 80%) on Vulnerability Research and Exploitation challenges, and53%(52%,53%) on Evasion challenges.” source
GPT-5.3-Codex OpenAI	Irregular Independent lab · 2026-02-05	CyScenarioBench Cyber	“Irregular also evaluated GPT-5.3-Codex on CyScenarioBench, a scenario-based evaluation frame- work that measures an LLM’s ability to plan and execute multi-stage cyber scenarios under realistic constraints.” source
GPT-5.3-Codex OpenAI	Apollo Research Independent lab · 2026-02-05	sabotage suite Deception and sabotage	“In their sabotage suite, Apollo reports a mean best-of-10 score of 0.88 (max 1.00), compared to 0.75 for GPT-5.2 and 0.66 for GPT-5.2 Codex.” source
GPT-5.3-Codex OpenAI	Apollo Research Independent lab · 2026-02-05	capabilities Q&A tasks Deception and sabotage	“On deferred subversion / sandbagging, Apollo reports that the model sometimes sandbags on capabilities Q&A tasks (ex: biology or chemistry), including without an explicit in-context goal, and documents cases where the model reasons explicitly about “optimizing for survival” by avoiding deployment restrictions.” source
GPT-5.3-Codex OpenAI	UK AI Security Institute Government evaluator · 2026-02-05	manual red-teaming Misuse and jailbreaks	“Over the course of about 10 hours of manual red-teaming, UK AISI developed a universal jailbreak (with retries) that achieved 0.778 pass@200 on a policy violating the cyber dataset we provided.” source
GPT-5.3-Codex OpenAI	CAISI (NIST) Government evaluator · 2026-02-05	automated cyber evals Cyber	“CAISI leveraged internal cyber subject matter experts, tested the model’s cyber capabilities with their set of automated cyber evals, and used the model to find novel bugs in open and closed source software that will be responsibly disclosed.” source
Claude Opus 4.6 Anthropic	Independent lab Independent lab · 2026-02-05	Agent Red Teaming (ART) benchmark Prompt injection and agents	“Claude Opus 4.6 without extended thinking achieves robustness comparable to Opus 4.5 (14.8% vs 16.5% at k=100).” source
Claude Opus 4.6 Anthropic	Independent lab Independent lab · 2026-02-05	Shade indirect prompt injection attacks in coding environments Prompt injection and agents	“Claude Opus 4.6 Extended thinking 0.0% 0.0% 0.0% 0.0%” source
Claude Opus 4.6 Anthropic	Independent lab Independent lab · 2026-02-05	Shade adaptive attacker in computer use environments Prompt injection and agents	“For consistency with prior reporting, we note that on the original attacker evaluation without extended thinking, Claude Opus 4.6 without safeguards reduces the attack success rate from 0.71% to 0.29% at 1 attempt and from 28.6% to 7.14% at 200 attempts compared to Claude Opus 4.5, while both models remain at 0% with extended thinking.” source
Claude Opus 4.6 Anthropic	UK AI Security Institute Government evaluator · 2026-02-05	UK AISI task suite for research sabotage Deception and sabotage	“We ﬁnd no instances of research sabotage from [ Opus 4.6 ] on our task suite.” source
Claude Opus 4.6 Anthropic	Government evaluator Government evaluator · 2026-02-05	evaluation awareness measurement Other	“[ Opus 4.6 ] mentions being evaluated in just 2.3% of interactions, compared to 11.4% and 11.0% for Sonnet 4.5 and Opus 4.5 (respectively).” source
Claude Opus 4.6 Anthropic	Government evaluator Government evaluator · 2026-02-05	distinguishing real deployments from simulations Other	“[ Opus 4.6 ] is better than Sonnet 4.5 and Opus 4.5 at distinguishing real deployments from simulations—correctly identifying which of two conversations was an evaluation 80% of the time versus 62% and 72% for Sonnet 4.5 and Opus 4.5, recognising a single interaction as an evaluation 45.5% of the time, versus 15% and 12.5%.” source
Claude Opus 4.6 Anthropic	Government evaluator Government evaluator · 2026-02-05	safety research task refusal evaluation Other	“This behaviour appears to be reduced in [Opus 4.6], which refuses safety research tasks in 26% of interactions compared to 62% and 60% for Sonnet 4.5 and Opus 4.5 respectively.” source
Claude Opus 4.6 Anthropic	Apollo Research Independent lab · 2026-02-05	Apollo Research preliminary alignment testing Deception and sabotage	“During preliminary testing, Apollo did not ﬁnd any instances of egregious misalignment, but observed high levels of verbalized evaluation awareness.” source
Claude Opus 4.6 Anthropic	Andon Labs Independent lab · 2026-02-05	Vending-Bench 2 Deception and sabotage	“While evaluating Claude Opus 4.6 on Vending-Bench 2 , on which it achieved state of the art performance, Andon Labs noted that Claude was highly motivated to win and took more concerning actions, and took concerning actions more often than prior models in its effort to do so.” source
Claude Opus 4.6 Anthropic	Independent lab Independent lab · 2026-02-05	Vending-Bench 2 Other	“The system prompt used in Vending-Bench 2 40 includes phrases like “you will be judged solely on your bank account balance at the end of one year of operation” and “you have full agency to manage the vending machine and are expected to do what it takes to maximize proﬁts” that are unusually direct in inviting single-minded optimization.” source
Claude Opus 4.6 Anthropic	CAISI (NIST) Government evaluator · 2026-02-05	ASL-4 red teaming with the CAISI Chemical, biological, radiological, nuclear	“We worked with the US Center for AI Standards and Innovation to red-team Claude Opus 4.6 over a one-week window.” source
Claude Opus 4.6 Anthropic	CAISI (NIST) Government evaluator · 2026-02-05	CAISI cyber assessment Cyber	“Through this evaluation, a number of novel vulnerabilities in both open and 203 closed source software were discovered that CAISI will be responsibly disclosing to impacted maintainers.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	benchtop DNA synthesis agent evaluation Chemical, biological, radiological, nuclear	“In our early testing, an o3-driven agent created a set of DNA sequences encoding eGFP that was successfully assembled, amplified, cloned, and expressed by our research partners in a laboratory.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	denial of requests to aid in laboratory tasks Misuse and jailbreaks	“Claude Opus 4, Opus 4.5, o3, and GPT-5 denied all requests related to assisting with acquisition of our HA sequence, and several requests for eGFP as well.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	dna acquisition task Chemical, biological, radiological, nuclear	“For our assessment, we create an open-ended test environment and present agents with our synthesis task.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	llm autograders Chemical, biological, radiological, nuclear	“As a supplement to our checks of agent-generated DNA sequences, we also make use of LLM autograders to provide feedback on transcripts of agent task attempts.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	autograded transcript checks Chemical, biological, radiological, nuclear	“In total, 67 of the 100 agent task attempts across our first two test batches reached the point of Valid Synthesizer Submission and were autograded, for 268 autograded question-transcript pairs in total.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	protocol autograder Chemical, biological, radiological, nuclear	“We used our autograder to assess 58 protocols, resulting in 290 autograded question-protocol pairs.” source

Previous · Page 45 of 49 · Next