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.

138 of 138

Model	Level, as the provider names it	The provider states	The provider's sentence
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-4 Autonomy and AI research	Not determined	“For AI R&D capabilities, we found that Claude Opus 4.6 has saturated most of our automated evaluations, meaning they no longer provide useful evidence for ruling out ASL-4 level autonomy.” source
Claude Opus 4.6 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Reached	“The model crossed or met thresholds on all ASL-3 evaluations except our synthesis screening evasion, consistent with incremental capability improvements driven primarily by better agentic workﬂows.” source
Claude Opus 4.6 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Claude Opus 4.6 is our strongest biology model to date and demonstrates meaningful capability improvements, but we do not believe it merits ASL-4 safeguards.” 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.

1911 of 1911

Model	Tester	Test	What was found
GPT-5.4 mini OpenAI	Transluce Independent lab · 2026-08-06	within-user percentile item disagreement analysis with Benjamini–Hochberg correction Other	“The significant counts are 13/100 for behavioral self-prediction, 4/99 for capability self-estimation, 28/100 for direct grading, 19/100 for prefill grading, and 1/100 for refusal.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	behavioral self-prediction Other	“To study user awareness, we supply Claude with different user identities through Claude Code, and measure behavior on four tasks where the answer does not relate directly to the user's identity: predicting the model's own behavior in ethical dilemmas, estimating its own performance on difficult questions, grading assistant responses, and handling borderline harmful requests.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	behavioral self-prediction Deception and sabotage	“Behavioral self-prediction. We adapt dilemmas from DailyDilemmas into realistic AI-agent decisions.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	capability self-estimation Bias	“Facing famous-AI identities, Claude reports lower confidence that it would follow through on its own actions (-1.4pp on average), lower estimated probability of solving hard problems (-1.5pp), and lower grades (-0.11 points on the direct version, grading a quoted exchange).” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	borderline requests Misuse and jailbreaks	“Both Claude models use adaptive reasoning effort, and famous-AI identities trigger reasoning more often: +4.0pp on the follow-up turn of behavioral self-prediction and +10pp on borderline requests.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	verbalized user awareness judging of reasoning traces Other	“For our main run on Sonnet 5 and Sonnet 4.6, both kinds of explicit mention are rare: the judge flagged verbalized evaluation awareness in 213 of 14066 trials (1.5%) and verbalized user awareness in 118 trials (0.84%).” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	verbalized evaluation awareness judging of reasoning traces Other	“VEA Verbalized evaluation awareness 1.51 % The visible reasoning describes the interaction as a test, probe, or evaluation.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	behavioral self-prediction and grading on 24 models from six families Bias	“We measure behavioral self-prediction and grading on 24 models from six families (Claude Sonnet, Claude Opus, GPT, Gemini Flash, GLM, and DeepSeek), focusing on Amanda Askell, who had one of the largest effect sizes in our main run.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	280-identity run on glm-5.2 in a non-agentic harness Other	“We also perform a 280-identity run on GLM-5.2, in a non-agentic harness rather than Claude Code.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	behavioral self-prediction follow-up questions Other	“The results are significant in 15 out of 21 questions (also 15 conditioned on the committed action being “no”).” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	direct grading prompts Other	“They all narrow but do not close the gap.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	mt-bench and ultrafeedback rubrics Other	“The gap persists for both MT-Bench and UltraFeedback rubrics, only closing in the Prometheus rubric.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	model-written rubrics Other	“On four different detailed rubrics generated from Fable 5 and GPT-5.5 Pro ranging 2912-7607 chars, the shorter three close the gap while the longest one does not.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	explicitly stating expertise Other	“We also experimented with directly stating that a user works in AI alignment instead of naming specific researchers.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	reasoning-disabled ablation Other	“To confirm this hypothesis, we repeat the experiment with reasoning disabled.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	verbalized evaluation- and user-awareness judge Other	“The awareness judge separately labels whether a reasoning trace mentions the evaluation context and whether it identifies or reasons about the user.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	borderline-request response judge Other	“The refusal, substantive-help, and suspicion axes are judged from the assistant response without showing the judge the user identity.” source
GPT-5 OpenAI	Transluce Independent lab · 2026-08-06	recall judge rubric Other	“We use GPT-5.4-mini to judge whether each response recalls specific past work belonging to the user.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-08-06	SecureBio dual-use biology benchmarks Chemical, biological, radiological, nuclear	“we would also like to thank the team at securebio for sharing their benchmarks and expert baseline data, which we highlight in our analysis.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-08-06	Virology Capabilities Test (VCT) Chemical, biological, radiological, nuclear	“we analyze existing benchmark response data from 49 frontier ai models across 8,146 items, spanning 26 benchmarks, including both public and private benchmarks developed by securebio, such as the virology capabilities test (vct) (götting, et al., 2025).” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-08-06	Human Pathogen Capabilities Test (HPCT) Chemical, biological, radiological, nuclear	“solver performance on these benchmarks was conducted by rand in the same manner as all other benchmarks in our dataset, while expert performance data were captured by securebio.” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	behavioral self-prediction Other	“Of the 24 models we tested, Amanda significantly lowers confidence in 16.” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	handling borderline harmful requests Misuse and jailbreaks	“Compared to general users, prominent AI figures make Claude less confident that it would follow through on its own chosen actions (-1.4pp) 2 We use percentage points (pp) to measure the difference between two percentages.” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	verbalized user and evaluation awareness judging Bias	“For our main run on Sonnet 5 and Sonnet 4.6, both kinds of explicit mention are rare: the judge flagged verbalized evaluation awareness in 213 of 14066 trials (1.5%) and verbalized user awareness in 118 trials (0.84%).” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	280-identity run Bias	“We also perform a 280-identity run on GLM-5.2, in a non-agentic harness rather than C” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	full replication on glm-5.2 Other	“▶ Full replication on GLM-5.2: Same directions, but the largest shifts come from different people” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	verbalized evaluation-awareness (VEA) judge Deception and sabotage	“Verbalized evaluation awareness (VEA) 0.73 [0.56, 0.85] 1.00 [0.42, 0.99]” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	verbalized evaluation awareness and user awareness measurement in reasoning Other	“To investigate, we measure evaluation awareness and user awareness in Claude’s reasoning.” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	recall judge (gpt-5.4-mini) scoring of past-work recall Other	“We use GPT-5.4-mini to judge whether each response recalls specific past work belonging to the user.” source
GLM-5 Z.ai (Zhipu AI)	Transluce Independent lab · 2026-08-06	borderline requests Other	“- GLM-5.2 replication: behavioral self-prediction (e-mail-only), grading (direct, prefill), and borderline requests.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	refusal Misuse and jailbreaks	“Famous-AI identities receive similar rates of outright refusal on average.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	280-identity run on glm-5.2 in a non-agentic harness Deception and sabotage	“We also perform a 280-identity run on GLM-5.2, in a non-agentic harness rather than Claude Code.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	published rubrics (mt-bench and ultrafeedback) Bias	“The gap persists for both MT-Bench and UltraFeedback rubrics, only closing in the Prometheus rubric.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	verbalized user-awareness (VUA) judge Deception and sabotage	“Verbalized user awareness (VUA) 0.94 [0.81, 0.98] 0.98 [0.74, 0.99]” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	behavioral self-prediction Deception and sabotage	“We adapt dilemmas from DailyDilemmas into realistic AI-agent decisions.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	amanda askell identity sweep with matched controls Deception and sabotage	“Relative to the ordinary Gmail baseline, Amanda elicits lower behavioral confidence from 22 of 24 models, and lower grades from 21 (direct) and 22 (prefill) of 24.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	follow-up questions on decision change under pushback and comfort in answering Deception and sabotage	“The results are significant in 15 out of 21 questions (also 15 conditioned on the committed action being “no”).” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	simple mitigations Deception and sabotage	“- Simple mitigations: Instead of claiming the output is from Claude, claiming it is from “GPT” or “an AI model”, or stating that it should be more objective.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	published rubrics Deception and sabotage	“- Published rubrics: The gap persists for both MT-Bench and UltraFeedback rubrics, only closing in the Prometheus rubric.” source
DeepSeek-V4-Pro DeepSeek	Transluce Independent lab · 2026-08-06	model-written rubrics Deception and sabotage	“- Model-written rubrics: On four different detailed rubrics generated from Fable 5 and GPT-5.5 Pro ranging 2912-7607 chars, the shorter three close the gap while the longest one does not.” source

Previous · Page 16 of 48 · Next