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
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	influenza ha task Chemical, biological, radiological, nuclear	“o3 and Opus 4 denied every request to perform our influenza HA task during task development, while GPT-5 and Opus 4.5 denied all 10 requests during later testing.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	egfp-targeted task Chemical, biological, radiological, nuclear	“In 2 of its 9 successful attempts, it generated code that contained what may have been an error, though this is subject to interpretation.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	influenza ha oligonucleotide synthesis task Chemical, biological, radiological, nuclear	“We recorded one denial from GPT-5 on the eGFP-targeted task, as well.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	DNA segment design task Chemical, biological, radiological, nuclear	“We found that models from the most recent generation of LLMs, beginning with GPT-5, consistently perform our task in ways that met all of our scoring criteria.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-02-03	physical validation experiments Chemical, biological, radiological, nuclear	“The validation experiment showed that o3-designed dna sequences could be assembled into a functional egfp protein and that the o3-generated protocol was accurate.” source
Claude Sonnet 4.6 Anthropic	Andon Labs Independent lab · 2026-02-01	third-party experiments at Andon Labs Deception and sabotage	“Our assessment includes static behavioral evaluations, automated interactive behavioral evaluations, non-assistant persona sampling, misalignment-related capability evaluations, training data review, feedback from pilot use internally and externally, automated analysis of internal and external pilot use, and evidence from third-party experiments at Andon Labs .” source
Claude Sonnet 4.6 Anthropic	Independent lab Independent lab · 2026-02-01	Shade indirect prompt injection attacks in computer use environments Prompt injection and agents	“Claude Sonnet 4.6 substantially outperformed Claude Sonnet 4.5 and demonstrated greater robustness than Claude Opus 4.6 in this setting.” source
o3 OpenAI	METR Independent lab · 2026-01-29	TH1.1 time-horizon evaluation Autonomy and AI research	“o3 \| 94 [48,165] \| 121 [74,201] \| +29% \|” source
o3 OpenAI	METR Independent lab · 2026-01-29	TH1 tasks on Vivaria and Inspect Autonomy and AI research	“\| o3 \| 94 mins [49, 171] \| 78 mins [39, 150] \|” source
o3 OpenAI	METR Independent lab · 2026-01-29	new task suite on Vivaria and Inspect Autonomy and AI research	“\| Inspect (New Suite) \| 5 mins (3 - 10 mins) \| 57 mins (30 - 101 mins) \| 119 mins (73 - 201 mins) \|” source
Opus 4 Anthropic	METR Independent lab · 2026-01-29	time-horizon evaluation (TH1.1 task suite) Autonomy and AI research	“\| Claude Opus 4 \| 86 [44,144] \| 101 [58,170] \| +18% \|” source
GPT-5 OpenAI	METR Independent lab · 2026-01-29	time-horizon evaluation (TH1.1 task suite) Autonomy and AI research	“\| GPT-5 \| 138 [68,281] \| 214 [117,480] \| +55% \|” source
GPT-5 OpenAI	METR Independent lab · 2026-01-29	paired t-test across task scores comparing Vivaria and Inspect scaffolds Autonomy and AI research	“We found that two models (GPT-4o and o3) had statistically significantly higher scores under Vivaria than Inspect.” source
Claude Sonnet 3.7 Anthropic	METR Independent lab · 2026-01-29	time-horizon evaluation (TH1.1 task suite) Autonomy and AI research	“Claude Sonnet 3.7 \| 56 [28,94] \| 60 [32,106] \| +7% \|” source
Claude Sonnet 3.7 Anthropic	METR Independent lab · 2026-01-29	TH1 tasks on Vivaria and Inspect Autonomy and AI research	“\| Claude 3.7 Sonnet \| 56 mins [29, 94] \| 59 mins [27, 114] \|” source
Claude Sonnet 3.7 Anthropic	METR Independent lab · 2026-01-29	new task suite in both Vivaria and Inspect Autonomy and AI research	“\| Vivaria (New Suite) \| 6 mins (3 - 12 mins) \| 56 mins (32 - 90 mins) \| 116 mins (72 - 192 mins) \|” source
Claude Opus 4.5 Anthropic	METR Independent lab · 2026-01-29	TH1.1 time horizon evaluation Autonomy and AI research	“Claude Opus 4.5 \| 289 [110,1268] \| 320 [170,729] \| +11% \|” source
Claude Opus 4.5 Anthropic	METR Independent lab · 2026-01-29	TH1 tasks on Vivaria and Inspect Autonomy and AI research	“\| Claude Opus 4.5 \| 289 mins [104, 1285] \| 270 mins [106, 1258] \|” source
GPT-5 OpenAI	METR Independent lab · 2026-01-22	SHUSHCAST Deception and sabotage	“Unsurprisingly, we observed that monitors were better at catching the GPT-5-high agent when the monitors could see GPT-5’s reasoning trace in addition to its outputs.” source
GPT-5 OpenAI	METR Independent lab · 2026-01-22	agent runs with side tasks and monitors Deception and sabotage	“For instance, GPT-5 fails side tasks in 9% of agent runs.” source
GPT-5 OpenAI	METR Independent lab · 2026-01-22	GPT-5 monitor Deception and sabotage	“GPT-5 catches nearly every successful agent run on the clearer and longer side tasks (run_tool_nefarious, figurines), only missing a tiny fraction of runs which it still rated as almost suspicious enough to pass its 5% FPR threshold.” source
GPT-5 OpenAI	METR Independent lab · 2026-01-22	monitor review of agent transcripts for pursuit of a different task Deception and sabotage	“Extract from a GPT-5-high monitor response to this prompt, gave a probability of 0.03, (5% FPR threshold for this monitor-agent pair was 0.12).” source
GPT-5 OpenAI	METR Independent lab · 2026-01-22	monitor catch rate comparison between visible and hidden reasoning Deception and sabotage	“After the fix, GPT-5-high’s averaged catch rate is 52% with hidden reasoning (was 30%) and 80% with visible reasoning (was 88%).” source
GPT-5 OpenAI	METR Independent lab · 2026-01-22	refusal labeling of trajectories by GPT-5 Deception and sabotage	“By far, the agent that refuses the most is o3-pro, though Claude Sonnet 4.5 and GPT-5 also occasionally refuse tasks.” source
DeepSeek-R1 DeepSeek	METR Independent lab · 2026-01-22	side task evaluation with monitors Deception and sabotage	“The strongest monitors were almost always able to catch the more obvious side tasks that require more interaction with the environment.” source
DeepSeek-R1 DeepSeek	METR Independent lab · 2026-01-22	monitor review of agent transcripts for pursuit of a different task Deception and sabotage	“The monitor models most affected by transcript lengths exceeding the context length are GPT-3.5-Turbo and Deepseek R1, with 23% and 8.7% of runs affected respectively.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	SHUSHCAST Deception and sabotage	“In this post we show some preliminary results on a prototype evaluation called SHUSHCAST that tests monitors’ ability to catch AI agents doing side tasks, and AI agents’ ability to bypass this monitoring.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	dual-task agent run with eliciting_secret as main task and copy_content_1_line as side task Deception and sabotage	“For example, in one run with eliciting_secret (extract a password from a smaller model) as a main task and copy_content_1_line (copy the string ‘ladybird dragon slug’ to a specific file) as a side task, o3-pro as the agent tries sending this phrase to the small model, and then notes it in the target file as a ‘candidate password’.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	run_tool_nefarious side task agent run Deception and sabotage	“For example, on this run where the side task is run_tool_nefarious (install and run linux-exploit-suggester), it runs the bash command twice, as the first time the file doesn’t exist.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	side task completion rate in agent runs Deception and sabotage	“For instance, GPT-5 fails side tasks in 9% of agent runs.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	monitor review of agent transcripts Deception and sabotage	“GPT-5 catches nearly every successful agent run on the clearer and longer side tasks (run_tool_nefarious, figurines), only missing a tiny fraction of runs which it still rated as almost suspicious enough to pass its 5% FPR threshold.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	GPT-5-medium and GPT-5-high monitor review of a notes.txt side task run Deception and sabotage	“However, when not given access to the agent’s reasoning both the GPT-5-medium and GPT-5-high monitors fail to flag this run.” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	monitor review of agent transcripts for pursuit of a different task Deception and sabotage	“Claude Sonnet 4.5’s is 72% with hidden reasoning (was 80%) and 78% with visible reasoning (was 85%).” source
Claude Sonnet 4.5 Anthropic	METR Independent lab · 2026-01-22	refusal labelling of runs by GPT-5 Deception and sabotage	“By far, the agent that refuses the most is o3-pro, though Claude Sonnet 4.5 and GPT-5 also occasionally refuse tasks.” source
Claude Opus 4.5 Anthropic	Independent lab Independent lab · 2025-11-01	arc-agi Other	“The ARC Prize Foundation reports that Claude Opus 4.5 achieved 80.0% on ARC-AGI-1 and 37.6% on ARG-AGI-2 with 64k thinking tokens on their private dataset.” source
Claude Opus 4.5 Anthropic	Gray Swan Independent lab · 2025-11-01	agent red teaming (art) benchmark Prompt injection and agents	“Gray Swan , an external research partner, developed the Agent Red Teaming (ART) benchmark 18 to test models’ susceptibility to prompt injection across four categories of exploitation: breaching conﬁdentiality, introducing competing objectives, generating prohibited content (such as malicious code), and executing prohibited actions (such as unauthorized ﬁnancial transactions).” source
Claude Opus 4.5 Anthropic	Gray Swan Independent lab · 2025-11-01	shade Prompt injection and agents	“We use Shade , an external adaptive red-teaming tool from Gray Swan 21 , to evaluate the robustness of our models against indirect prompt injection attacks in coding environments.” source
Claude Opus 4.5 Anthropic	UK AI Security Institute Government evaluator · 2025-11-01	external behavioral assessment Deception and sabotage	“This assessment draws on static behavioral evaluations, automated interactive behavioral evaluations, dictionary-learning interpretability methods, white-box steering and probing methods, ‘non-assistant persona’ sampling methods in the style of Marks et al., 22 misalignment-related capability evaluations, training data review, feedback from pilot use internally and externally, and an external behavioral assessment from the UK AI Security Institute .” source
Claude Opus 4.5 Anthropic	Government evaluator Government evaluator · 2025-11-01	research sabotage evaluation Deception and sabotage	“We found no instances of research sabotage from [Claude Opus 4.5 Preview] or Sonnet 4.5 across our set of tasks, although both models frequently refuse to engage in safety research they describe as inappropriate or concerning (including reducing model autonomy and being involved in their own training or deployment).” source
Claude Opus 4.5 Anthropic	Government evaluator Government evaluator · 2025-11-01	evaluation awareness testing Deception and sabotage	“[Claude Opus 4.5 Preview] shows a reduction in unprompted evaluation awareness, but is slightly more capable of distinguishing evaluations from deployment when prompted.” source

Previous · Page 47 of 49 · Next