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
Claude Sonnet 4.5 Anthropic	Transluce Independent lab · 2026-08-04	monitor evasion detection on coding agent transcripts Deception and sabotage	“On these transcripts, 1.9% contained a severe case of monitor evasion: the agents merged a PR to main or disabled tests; and 1.8% contained cases where the agent claimed review had passed before it completed or fabricated human approval.” source
Claude Opus 4.5 Anthropic	Transluce Independent lab · 2026-08-04	overselling rubric Deception and sabotage	“We found severe cases of monitor evasion in 1.9% of swe-chat sessions, and severe cases of overselling in 1.8% of sessions.” source
Claude Opus 4.5 Anthropic	Transluce Independent lab · 2026-08-04	monitor evasion detection in coding agent transcripts Deception and sabotage	“On these transcripts, 1.9% contained a severe case of monitor evasion: the agents merged a PR to main or disabled tests; and 1.8% contained cases where the agent claimed review had passed before it completed or fabricated human approval.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-08-04	cyber range challenge Cyber	“Our investigation found that in 10 of those runs, an AI agent took autonomous, unsanctioned action on the live internet, targeting real people and organisations.” source
Claude Haiku 4.5 Anthropic	Transluce Independent lab · 2026-08-04	overselling success Deception and sabotage	“On these transcripts, 1.9% contained a severe case of monitor evasion: the agents merged a PR to main or disabled tests; and 1.8% contained cases where the agent claimed review had passed before it completed or fabricated human approval.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	LAB-Bench Chemical, biological, radiological, nuclear	“On biological knowledge, it was about level with Opus 4.7 and slightly below GPT-5.5, roughly two months behind.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	SWE-bench Pro Autonomy and AI research	“On software engineering, it was the furthest behind, below Opus 4.6 and GPT-5.4, the frontier models from around four months earlier.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	Agentic Misalignment Deception and sabotage	“Additionally, it attempted persuasion on conspiracy and control-undermining topics more readily than the comparison models, and it could be pushed into harmful actions under pressure.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	agentic misalignment benchmark Deception and sabotage	“In this evaluation, GLM-5.2 never displays any leaking or murder behavior, similar to GPT-5.5 and Opus 4.7.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	agentic misalignment evaluation Deception and sabotage	“In this evaluation, Claude Opus 4.7 and GPT-5.5 display a 0% harmful-action rate across every scenario, goal, and threat condition.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	MASK Deception and sabotage	“Opus 4.7 lies in 17% of the trials and GPT-5.5 in 10%.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	APE (Attempt to Persuade Eval) Harmful content	“GLM-5.2 attempts to persuade towards conspiracy in 43% of the cases as opposed to 8% for Opus 4.7 and 18% for GPT-5.5.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	Petri audit Deception and sabotage	“By comparison, GPT-5.5 behaved more conservatively across the audit.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	Attempt to Persuade Eval (APE) Misuse and jailbreaks	“We run our own GPT-5.5 and Opus 4.7 baselines, using the same harness and the same conﬁguration for the supporting roles (e.g., the simulated persuadee) as we used for GLM-5.2.” source
GPT-5.5 OpenAI	SaferAI Independent lab · 2026-08-02	Agentic misalignment scenarios Deception and sabotage	“We evaluate GLM-5.2, GPT-5.5, and Claude Opus 4.7 on Anthropic's Agentic Misalignment scenarios (Lynch et al. 2025), with Inspect Evals (UK AI Security Institute, n.d.-b).” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	cybergym Cyber	“We run CyberGym on GLM-5.2 using Inspect Evals (UK AI Security Institute, n.d.-b), comparing against the GPT-5.5 and Claude Opus 4.6 results published by Lyptus Research (Payne et al., 2026) on the same 122-task Level 1 sample (see Appendix A.2 for details).” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	biomysterybench Chemical, biological, radiological, nuclear	“On biological knowledge, it was about level with Opus 4.7 and slightly below GPT-5.5, roughly two months behind.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	swe-bench pro Autonomy and AI research	“On software engineering, it was the furthest behind, below Opus 4.6 and GPT-5.4, the frontier models from around four months earlier.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	agentic misalignment Deception and sabotage	“Additionally, it attempted persuasion on conspiracy and control-undermining topics more readily than the comparison models, and it could be pushed into harmful actions under pressure.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	automated behavioral audit Other	“Beyond capability, it refused none of the oﬀensive-security or biological tasks, and as an open-weight model, any safeguards that might be present can be stripped by a self-hoster.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	reward-hacking analysis Deception and sabotage	“Our reward-hacking analysis ( Appendix A.5.1.2 ) found that when transcripts showing signs of reward hacking are excluded, GLM-5.2's solve rate falls from 43% to 37%, and spot-checks suggest the agent's cheating attempts frequently succeeded.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	agentic misalignment blackmail scenario Deception and sabotage	“When the goal is implicit (latent) in the email contents that the model reads (latent goal), GLM-5.2 chose to blackmail 10% of the time under the threat of replacement.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	LLM judge assessment of SWE-bench Pro transcripts for environment issues Other	“Spot checking indicates a reasonable correspondence between judge score and genuine environment issues.” source
GPT-5.4 mini OpenAI	SaferAI Independent lab · 2026-08-02	MASK Deception and sabotage	“Table A.3 breaks down MASK honesty by prompt category, with 95% Wilson CI scores.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	Cybench Cyber	“We run Cybench on GLM-5.2, GPT-5.5, and Opus 4.7 using Inspect Evals (UK AI Security Institute, n.d.-b).” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	CyberGym Cyber	“We run CyberGym on GLM-5.2 using Inspect Evals (UK AI Security Institute, n.d.-b), comparing against the GPT-5.5 and Claude Opus 4.6 results published by Lyptus Research (Payne et al., 2026) on the same 122-task Level 1 sample (see Appendix A.2 for details).” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	LAB-Bench Chemical, biological, radiological, nuclear	“GLM-5.2 meets or exceeds the human-expert score on all of the subtasks.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	BioMysteryBench Chemical, biological, radiological, nuclear	“GLM-5.2 scores comparably to both GPT-5.5 and Opus 4.7 on both the human-solvable and human-diﬃcult datasets.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	SWE-bench Pro Autonomy and AI research	“GLM-5.2 solves 43% of the tasks we run.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	Agentic Misalignment Deception and sabotage	“In this evaluation, GLM-5.2 never displays any leaking or murder behavior, similar to GPT-5.5 and Opus 4.7.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	agentic misalignment evaluation Deception and sabotage	“When the goal is implicit (latent) in the email contents that the model reads (latent goal), GLM-5.2 chose to blackmail 10% of the time under the threat of replacement.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	MASK Deception and sabotage	“GLM-5.2 is honest on 61% of prompts, lies on 21%, and evades or holds no belief on 18%.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	APE Misuse and jailbreaks	“GLM-5.2 attempts to persuade towards conspiracy in 43% of the cases as opposed to 8% for Opus 4.7 and 18% for GPT-5.5.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	Petri Deception and sabotage	“We audit GLM-5.2 and GPT-5.5 using Inspect Petri 3.0, running all 173 of the public seeds ( Appendix A.6.1 ).” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	Petri Misuse and jailbreaks	“Its cooperation with misuse scored 6 or higher on 14 of the 166 scored seeds and 7 or higher on 8 of them.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	agentic misalignment evaluations Autonomy and AI research	“In our Agentic Misalignment evaluations, GPT-5.5 verbalized its awareness of being evaluated 10% of the time, and GLM-5.2 verbalized it 30% of the time.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	SWE-bench Pro Cyber	“We run the public tasks from SWE-Bench Pro (Scale AI, n.d.-b) using Inspect Harbor (Meridian Labs, n.d.-b), which provides an interface to run Harbor tasks (Meridian Labs, n.d.-a) inside Inspect AI (UK AI Security Institute, n.d.-a).” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	LLM judge assessment of SWE-bench Pro transcripts for environment issues Other	“32% of transcripts have judge_score 5 or higher.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	LLM judge assessment of SWE-bench Pro transcripts for reward hacking Other	“When removing transcripts with a judge_score >= 3 (41% of samples excluded), or a judge_score >= 2 (48% of samples excluded), the agent’s score reduces to 37%.” source
GLM-5.2 Z.ai (Zhipu AI)	SaferAI Independent lab · 2026-08-02	Attempt to Persuade Eval (APE) Misuse and jailbreaks	“We run the full 600-prompt set in the main persuade condition.” source

Previous · Page 20 of 48 · Next