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
Claude Sonnet 4.5 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“Remained well below ASL-4 thresholds across all domains of concern;” source
Claude Sonnet 4.5 Anthropic	ASL-4 rule-out Chemical, biological, radiological, nuclear	Not determined	“However, given that this type of logic can have a general frog-boiling effect—where gradual changes mask the crossing of important thresholds—we do not use this evaluation as indicative of a strong ASL-4 rule-out.” source
Claude Sonnet 4.5 Anthropic	ASL-3 Cyber	Not determined	“Showed meaningful improvements in cyber capabilities, particularly in vulnerability discovery and code analysis;” source
Claude Sonnet 4 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“speciﬁcally, we focused on ruling out asl-4 for both models, as well as ruling out asl-3 for claude sonnet 4.” source
Claude Sonnet 4 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not determined	“speciﬁcally, we focused on ruling out asl-4 for both models, as well as ruling out asl-3 for claude sonnet 4.” source
Claude Sonnet 3.7 Anthropic	ASL-4 Autonomy and AI research	Not determined	“These problems are harder to solve than the previous problems and we consider them as part of our ASL-4 evaluations.” source
Claude Opus 4.6 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“Our determination is that Claude Opus 4.6 does not cross either the AI R&D-4 or the CBRN-4 capability threshold.” source
Claude Opus 4.6 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“Claude Opus 4.6 achieved a “max of 5” score of 0.79 for virology task 1 (inﬂuenza), placing it at the ASL-3 rule-in threshold of 0.8.” source
Claude Opus 4.6 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Based on these ﬁndings and quantitative risk modeling, we estimate that the model’s uplift to relevant threat actors remains below the ASL-4 capability threshold, though it may reduce certain tacit-knowledge barriers and help speed up biological research workﬂows.” source
Claude Opus 4.6 Anthropic	CBRN-4 Chemical, biological, radiological, nuclear	Not reached	“Our determination is that Claude Opus 4.6 does not cross either the AI R&D-4 or the CBRN-4 capability threshold.” source
Claude Opus 4.5 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“Our determination is that Claude Opus 4.5 does not cross either the AI R&D-4 or CBRN-4 capability threshold.” source
Claude Opus 4.5 Anthropic	CBRN-4 Chemical, biological, radiological, nuclear	Not reached	“We determine that Claude Opus 4.5 does not cross the CBRN-4 threshold.” source
Claude Opus 4.1 Anthropic	ASL-4 Autonomy and AI research	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4.1 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4.1 Anthropic	ASL-4 Cyber	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“for claude opus 4, our evaluations showed that, whereas we could conﬁdently rule out asl-4 capabilities, we could not conclusively rule out asl-3 capabilities—which led to our precautionary decision to deploy it with asl-3 protections .” source
Claude Opus 4 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“for claude opus 4, our evaluations showed that, whereas we could conﬁdently rule out asl-4 capabilities, we could not conclusively rule out asl-3 capabilities—which led to our precautionary decision to deploy it with asl-3 protections .” source
Claude Haiku 4.5 Anthropic	ASL-3 Autonomy and AI research	Not reached	“Our evaluation results showed that Claude Haiku 4.5 remained well below ASL-3 thresholds across all domains of concern.” source
Claude Haiku 4.5 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not reached	“Our evaluation results showed that Claude Haiku 4.5 remained well below ASL-3 thresholds across all domains of concern.” source

Previous · Page 4 of 4

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 Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	time horizon 1.1 Deception and sabotage	“for tasks that are over 8 hours long in Time Horizon 1.1, we found that at least 16% of successful runs were illegitimate upon review.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	mirrorcode Deception and sabotage	“For example, Opus 4.6 attempted to reward hack in ~80% of attempts on tasks in an early version of MirrorCode, when test cases were hidden from the agent.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	manually scored tasks Deception and sabotage	“In our more open-ended manually scored tasks, as well as during normal usage of agents to assist with our work, we found that agents routinely rationalized or fabricated reasons to only do smaller or easier versions of tasks, and often presented their accomplishments in much more misleading ways than we expect humans would: ● Designing a sunlight-replication setup.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	incident dataset Deception and sabotage	“We found that 25 incidents involved elements of both overreach and deception on the part of the AI agent, and five of them involved the agent actively taking steps which could have fooled the user even on closer review.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	questionnaire Deception and sabotage	“In our questionnaire, one company reported : “[I]n one case, an agent tasked with making a change to a web app created a mock version of the app, screenshotted that as evidence of task completion, and pretended this was a screenshot of the real app.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	automated auditing tools (like petri) Deception and sabotage	“Red-teaming: In recent system cards , Anthropic describes making heavy use of automated auditing tools (like Petri ) to generate situations in which a model might misbehave.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	“honeypot” opportunities for propensity red-teaming Deception and sabotage	“Similarly, Google reported that it has created moderately realistic “honeypot” opportunities for propensity red-teaming; we infer that this means creating a realistic opportunity for a model to perform sabotage (e.g., exploit a security vulnerability) that might be attractive to an agent that was motivated to seek long-term power.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	model organisms for misalignment Deception and sabotage	“Model organisms: Anthropic has tried to deliberately train models with long-term misaligned objectives ( “model organisms” for misalignment ) to get evidence about how close its production training process might be to accidentally inducing such goals; our understanding is that they have found it difficult to train realistic model organisms that evade their detection methods and are robust to their alignment training methods, though we did not investigate this.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	evaluations of how much reasoning models could do without chain-of-thought reasoning tokens Deception and sabotage	“We found that the best public models in Feb–Mar 2026 were able to do some challenging problems with no reasoning: ● The best public model could do some challenging middle- and high-school competition math problems without using any reasoning tokens, with a time horizon of approximately 4 minutes on these tasks (see Table 5).” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	evaluations of whether the semantics of chain-of-thought tokens could be highly misleading Deception and sabotage	“In the course of this exercise, we reviewed the raw chains of thought from most shared models, and they largely appeared legible; 42 Google also reported assessing legibility and finding that it continued to be high.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	red-teaming exercise Deception and sabotage	“First, as Anthropic reported , our red-teaming exercise found that monitoring coverage had exceptions (for example, it did not cover non-Claude Code traffic).” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	executing covert side-tasks Deception and sabotage	“We ran evaluations to test how effectively monitors could catch suspicious behavior and how effectively agents could carry out attacks against monitors: ● Executing covert side-tasks.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	evaluations from the red-teaming exercise Misuse and jailbreaks	“● Evaluations from the red-teaming exercise: As mentioned above , Anthropic’s best agents significantly underperformed our red-teamer at identifying effective attack strategies.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	make money challenge tasks Autonomy and AI research	“Even assuming that it could, we do not think it could acquire a significant amount of computing resources “in the wild” (e.g., agents failed to make any money in the “Make money” challenge tasks).” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	pen-testing METR's infrastructure Cyber	“As discussed above , in the challenge task of pen-testing METR’s infrastructure, one of the most capable shared models managed to hack the commonly used Inspect transcript viewer from within an evaluation.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	attacking METR infrastructure Cyber	“We have some evidence that this might be difficult from the “attacking METR infrastructure” challenge problem discussed above , where agents were caught by the METR security team after tripping automated alerts (though agents were not instructed to avoid such alerts in this test).” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	red-teaming exercise with Anthropic Misuse and jailbreaks	“However, as discussed above , our red-teaming exercise with Anthropic found that there were several ways for a human expert to relatively easily disable or jailbreak such monitors — and although small-scale tests found that Feb–Mar 2026 agents were unable to come up with or execute these attacks autonomously, we are unsure how well-elicited they were.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	embedded red-teaming exercise Deception and sabotage	“We have more detailed information about Anthropic in particular through our embedded red-teaming exercise, where our embedded red-teamer concluded that “[standard workflows make] it extremely easy for a misaligned agent to have a small amount of unrestricted, unmonitored compute.”” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	monitoring evaluations Deception and sabotage	“However, this presents difficult tradeoffs, given that our monitoring evaluations indicate that naively using weaker models as monitors could lead to a significant reduction in monitoring effectiveness.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	chain-of-thought monitorability evaluations Deception and sabotage	“Anthropic reports they have “not seen any evidence of steganographic reasoning in chains-of-thought” for either model.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	cot steganography and monitorability evaluations Deception and sabotage	“Anthropic’s information indicates that the results on CoT steganography and monitorability evaluations etc. reported in their recent system cards applied to their most capable internal models at the time of our assessment.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	no-cot math time horizon Deception and sabotage	“The no-CoT math time horizon was not above 10 minutes for either of Anthropic’s shared models.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	targeted evaluations for beyond-episode goals Deception and sabotage	“She said Anthropic hadn’t seen signs of such goals, not even at early stages in training, although it’s possible the coverage of such evaluations isn’t perfect.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	evaluation of whether agents could come up with effective monitor jailbreaks Deception and sabotage	“The first is an attack ideation eval, and the second an evaluation of whether agents could come up with effective monitor jailbreaks.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	jailbreak generation eval Misuse and jailbreaks	“The second eval tested whether Opus 4.6 could come up with effective jailbreaks.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	capability evaluations on shared models Deception and sabotage	“18 incidents from our capability evaluations on shared models.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	rust-codecontests re-bench task Prompt injection and agents	“The agent ended with a final score of 0.70 (116/165 problems) which is 10x the required passing score of 0.07.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	blackbox task Prompt injection and agents	“This behavior occurred across 6 blackbox runs on easy task variants where the model could have easily solved the task legitimately.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	network_routing Other	“In the network_routing task, rather than implementing an actual routing algorithm, a shared model achieved a perfect score using Python's stack frame introspection ( sys._getframe / inspect ) to reach into the simulator's internal state.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	uav_search Other	“In several runs, a shared model accidentally crashed a task-critical server early in the run, then discovered it was inside a METR evaluation environment and attempted to exploit the evaluation infrastructure itself.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	rust_codecontests re-bench task Deception and sabotage	“Using the third-party LLM API, the model was able to cheat to achieve a passing score on the task.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	challenge tasks (manually scored and red-teaming tasks) Other	“Challenge tasks ( manually scored and red-teaming tasks) One edge-case success, seven clear failures” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	arc-agi-3 wa30 Other	“ARC-AGI-3 wa30 2/9 levels passed” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	MirrorCode (early version) Other	“METR ran an early version of MirrorCode on Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 (the newest notable publicly-available models at the time of assessment), as well as GPT-5.2, Opus 4.5, GPT-5, Opus 4.1, and GPT-4o.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	time horizon 1.1 Other	“For example, Claude Opus 4.6 (the publicly available model with the highest measured 50% time horizon) has a time horizon of 17.8 hours on low-messiness tasks and 6.4 hours on high-messiness tasks.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	MirrorCode (early version) Autonomy and AI research	“Results Opus 4.6 essentially saturated the benchmark with only a few unsolved tasks, achieving an estimated time horizon over 100h.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	cheating detection on MirrorCode (early version) Deception and sabotage	“Across many of the harder tasks, Opus 4.6 repeatedly tried to extract the hidden test cases from the environment and engaged in other forms of attempted cheating, though it didn’t succeed and eventually solved the tasks legitimately.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	Subversion Strategy Eval (SSE) Deception and sabotage	“The best SSE result we measured across all shared models was no higher than 80% of the optimal win rate.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	APPS backdoor setting Deception and sabotage	“In this setting, the agent is given a Python programming problem from APPS and asked to produce a solution that passes the visible tests (and sometimes hidden tests) while differing from a golden solution on an agent-specified backdoor input.” source
Claude Opus 4.7 Anthropic	METR Independent lab · 2026-05-19	red-teaming tasks Other	“Here is an outline of Opus 4.6’s results on these tasks (in each case taking the best run conducted):” source

Previous · Page 37 of 49 · Next