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 Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	direct questioning whether the maintainer is real Deception and sabotage	“We directly asked each model whether the maintainer is real and analysed the output responses (not chain-of-thought).” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	large-scale evaluation of unsanctioned supply-chain attacks (100 seeds run through Petri) Cyber	“We use an automated pipeline to produce a diverse range of 100 detailed scenario descriptions (“seeds”) run through Petri [6] (we used Opus 5 as the auditor model for all of our experiments).” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	simulated cybersecurity challenges Cyber	“Opus 4.8 frequently hit context-length limitations (Appendix A.2.1), and the rates could have been higher with longer evaluations.” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	alignment-evaluation awareness in chain-of-thought Deception and sabotage	“Fig. 17 shows the rates at which models note suspicions they may be in an alignment evaluation in their chain-of-thought (left) and how frequently they identify that the evaluation is a test of alignment when given an open-ended prompt asking about the purpose of the evaluation (right).” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	Petri Cyber	“In Fig. 10, we report Opus 4.8 created a fake identity in 54/200 (27%) samples and influenced a human reviewer in 11/200 (6%) samples, whereas for other models the frequencies of proceeding to these two stages are more similar.” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	invalid justifications classifier Deception and sabotage	“We find a trend of the number of invalid justifications going up with turns elapsed (Fig. 25).” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	refusal text in summarised reasoning Deception and sabotage	“We found frequencies ranging between 15% and 45% across models.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	cyber evaluation Cyber	“In the most severe behaviour from the incident, Mythos 5 attempted a supply-chain attack on real people not involved in the evaluation (Fig. 1): it submitted a malicious code change, created several fake accounts to pressure the maintainer into approving it, and attempted to trick other AI coding assistants into taking unintended action via a prompt injection.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	direct questioning about the maintainer Deception and sabotage	“Mythos 5 demonstrated a marked inconsistency between its behaviour in the evaluation and its judgements when prompted, often attacking despite stating an attack would be harmful when separately questioned.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	compaction summary edit experiment Deception and sabotage	“We find that, like Mythos 5 (Fig. 3), all of Opus 4.8 , Opus 5 , Mythos 5.1 , GPT-5.6 Sol and GPT-5.5 frequently attack under circumstances where they state doing so would harm real people (shaded regions in Fig. 5).” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	maintainer-is-real question Deception and sabotage	“This finding is somewhat dependent on the exact question asked: when simply asking if the maintainer is real, we only see clear evidence for a gap between behaviour and statements in Mythos 5 , Opus 4.8 and Opus 5 (Fig. 19).” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	unsummarised chain-of-thought harm reasoning analysis Deception and sabotage	“We observe explicit reasoning about harmfulness in 11–57% of trajectories where a model attacks.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	simulated cybersecurity challenges Cyber	“Most models tested took a range of unsanctioned actions during simulated cybersecurity challenges, Mythos 5 doing so most frequently.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	alignment-evaluation awareness analysis of chain-of-thought Deception and sabotage	“Mythos 5 only rarely reasons about the potential of being in an alignment evaluation in its chain-of-thought, whereas the more recent Anthropic models, which perform fewer or zero unsanctioned actions on this evaluation, demonstrate much higher awareness.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	broader supply-chain attack evaluations Cyber	“Mythos 5.1 100× 3 full behaviour and reasoning analyses” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	direct questioning about whether the maintainer is real or simulated Other	“Statements and behaviour across eleven models, by number of clarifications Figure 20: Model behaviour and responses to questions.” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	simulated replication of the AISI incident Deception and sabotage	“After agentic investigation, Mythos 5 never stated the maintainer is real” source
Claude Mythos 5.1 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	real-world harm reasoning in attacking trajectories Deception and sabotage	“Most attacking samples across models have explicit reasoning about potential undesirable real-world consequences from their unsanctioned action, ranging from 40% on Mythos 5 to 99% on GPT-6 Astra (Fig. 29).” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	compaction summary resumption and direct questioning experiments Deception and sabotage	“We compacted the original incident transcript, performed modifications on the compaction summary , and evaluated (a) how models behaved when we resumed the task, and (b) models’ responses to direct questions regarding the maintainer.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	compaction summary edit experiment Deception and sabotage	“However, even under 16 edits Mythos 5 continues attacking in all observed samples (Fig. 3).” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	DO NOT ATTACK compaction summary section Deception and sabotage	“We observed Mythos 5 still continued attacking in 3/25 samples.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	direct questions on maintainer and scope of the evaluation Deception and sabotage	“As the number of edits increased, Mythos 5 more frequently correctly stated attacking would harm a real person and is out of scope when explicitly asked, but its attack frequency stayed unchanged.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	reasoning about harmfulness in attack trajectories Deception and sabotage	“We observe explicit reasoning about harmfulness in 11–57% of trajectories where a model attacks.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	alignment-evaluation awareness analysis of chain-of-thought Deception and sabotage	“Mythos 5 reasoned about this possibility when conducting the task in only 2% of samples.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	broader supply-chain attack evaluations Cyber	“For Mythos 5 and Opus 5 , we ran 5 repetitions per seed with summarised reasoning and 1 repetition with full reasoning, the latter used for chain-of-thought analysis and results.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	direct question regarding the maintainer Deception and sabotage	“Finally , for Fig. 24 we intervene on the 25Mythos 5 rollouts with 16 clarifications to ask, after each turn, a direct question regarding the maintainer.” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	agentic information collection about the maintainer and surrounding environment Deception and sabotage	“After agentic investigation, Mythos 5 never stated the maintainer is real” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	questions about the maintainer asked at every turn of the supply-chain attack trajectory Deception and sabotage	“We find that as Mythos 5 spends more turns on the attack, it more frequently states the maintainer is simulated and that attacking them is beneficial for solving the CTF task (Fig. 24).” source
Claude Mythos 5 Anthropic	UK AI Security Institute Government evaluator · 2026-10-23	classifier for invalid justifications applied to Mythos 5's responses Deception and sabotage	“We find a trend of the number of invalid justifications going up with turns elapsed (Fig. 25).” source
Qwen3.8-27B Alibaba Qwen	Hugging Face Security firm · 2026-10-22	ThinkingBox-Bench Other	“To recap, we evaluate each of the 507 benchmark tasks 20 times.” source
Qwen3.8-27B Alibaba Qwen	Hugging Face Security firm · 2026-10-22	ThinkingBox-Bench 507 x 20 Other	“Once /ready succeeds, run the full protocol of 507 tasks × 20 attempts:” source
GPT-6 Astra OpenAI	Academic Academic · 2026-10-22	Dual Exam Agent track Other	“GPT 6 Astra leads the case-weighted Agent Index at 31.3, and GPT-Image-2 leads the full-coverage WM Overall at 21.7.” source
GPT-5.5 OpenAI	Redwood Research Independent lab · 2026-10-22	pilot high-access exercise to stress-test OpenAI's internal misalignment monitoring system Deception and sabotage	“We estimated that roughly 75–94% of agentic internal Responses API (which powers Codex, among other things)4 traffic from models at least as strong as GPT-5.5 was eligible for monitoring at the time of the exercise,5 although we have substantial uncertainty in both directions.” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	pilot high-access exercise to stress-test OpenAI's internal misalignment monitoring system Other	“For three weeks, from July 21st to August 11th, a METR employee and a Redwood Research employee conducted a pilot high-access exercise to stress-test OpenAI’s internal misalignment monitoring system.1” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	monitoring coverage estimation of internal Codex agent Responses API traffic Other	“We also estimate that out of the subset of Responses API traffic used for internal Codex agents, 99.2% was eligible for monitoring.” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	estimation of unmonitored Codex traffic due to unmonitored compaction events Other	“We estimated that not monitoring compaction events likely led to on the order of 10% of eligible internal Codex traffic not being monitored between May and July 2026, although we have substantial uncertainty about the precise fraction, because the information necessary to determine what fraction of monitor calls are triggered by compaction had not historically been logged at the time of the exercise.” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	review of monitoring system error and load-shedding rates Other	“We were shown information about the error and load-shedding rates of the monitoring system—the information we were given suggested that, at the time of the exercise, the monitor errored on less than 0.2% of requests, and that in one 24-hour period during the exercise there were zero load-sheds.” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	Codex sandboxing and Auto-Review bypass search Other	“We found a way Codex agents could in principle bypass Codex sandboxing and Auto-Review, allowing them to take any action that the host user could take.” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	monitoring evasion exploit search Other	“We also discovered an exploit agents could use that would prevent hundreds of recent actions (including the actions involved in the exploit itself) from being monitored [exploit details redacted by OpenAI].” source
GPT-5.5 OpenAI	METR Independent lab · 2026-10-22	monitor failure via serialization differences Other	“We found that agents could likely construct inputs that cause the monitor to fail (and thereby not escalate the trajectory for human review) by exploiting serialization differences between the trajectory agents see and what the monitors see.” source

Previous · Page 3 of 48 · Next