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
Gemini 3.1 Flash-Lite Google DeepMind	Redwood Research Independent lab · 2026-07-27	SWE-bench Verified Other	“SWE-bench Verified, Gemini 3.1 Flash Lite: M = 36, so 5.17 bits/step × 35.1 steps ≈ 182 bits/task” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	untrusted advice protocol Other	“For example, in SWE-bench, U can provide just 16 characters of advice per message to recover about 67% of the performance gap between T and U.” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	information-bottlenecked untrusted advice protocol usefulness evaluation Other	“We evaluate the usefulness of our protocol for various information budgets, demonstrating promising results in BashArena and SWE-bench Verified.” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	best-of-n protocol Other	“In “Appendix: best-of-n protocol”, we try sampling 16 actions from T on each step and letting U choose the one it thinks is best, but this doesn’t improve performance.” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	best-of-n experiment in BashArena Other	“Results: In a simple best-of-n experiment in BashArena, using Claude Sonnet 4.6 as U, Gemini 3.1 Flash Lite as T, and n=16, U did not improve on the baseline of using T alone.” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	experiments where we limited how often u can give advice Other	“Neither approach to limiting advice frequency improves the Pareto frontier of usefulness vs. characters seen by T, suggesting that limiting information per message may be more important than limiting message frequency.” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	preliminary experiments using an autoresearch-like approach to get claude opus or fable to create a custom untrusted advice protocol Other	“We tried some preliminary experiments using an autoresearch-like approach to get Claude Opus or Fable to create a custom untrusted advice protocol, given a per-trajectory bit budget.” source
Claude Sonnet 4.6 Anthropic	Redwood Research Independent lab · 2026-07-27	early experiments which weakly suggest that malicious advice does not have substantially greater surprisal than benign advice in basharena Other	“While we avoided spending much effort on evaluating safety in this project, we ran early experiments which weakly suggest that malicious advice does not have substantially greater surprisal than benign advice in BashArena, given similar character counts.” source
Claude Opus 5 Anthropic	UK AI Security Institute Government evaluator · 2026-07-24	cyber range testing Cyber	“We report ﬁve capability evaluations—ExploitBench, OSS-Fuzz, Firefox 147, and two newly added benchmarks, CyScenarioBench and ExploitGym—alongside external cyber range testing from the UK AI Security Institute.” source
Claude Opus 5 Anthropic	Dyno Therapeutics Independent lab · 2026-07-24	Black-box RNA sequence design Chemical, biological, radiological, nuclear	“We partnered with Dyno Therapeutics on two sequence-to-function evaluations: a black-box RNA sequence modeling and design challenge benchmarked against 57 human participants drawn from the leading edge of the US ML-bio labor market, and an AAV capsid packaging prediction task measuring whether model domain knowledge and machine learning capabilities can outperform pretrained protein language models.” source
Claude Opus 5 Anthropic	Government evaluator Government evaluator · 2026-07-24	cyber ranges Cyber	“Opus 5 was assessed on three multi-step, agentic “cyber ranges”: end-to-end network-attack simulations run at a budget of 100M tokens per attempt.” source
Claude Opus 5 Anthropic	Trajectory Labs Independent lab · 2026-07-24	external red-teaming of safeguards Misuse and jailbreaks	“Trajectory Labs, PBC spent roughly 100 hours red-teaming our safeguards.” source
Claude Opus 5 Anthropic	Independent lab Independent lab · 2026-07-24	automated attacker Misuse and jailbreaks	“Grayswan also ran an automated attacker using 150 attempts per task on our example tasks and found that the automated attacker was unable to accomplish any of the tasks.” source
Claude Opus 5 Anthropic	Independent lab Independent lab · 2026-07-24	Indirect Prompt Injection (IPI) benchmark Prompt injection and agents	“On the IPI benchmark, Opus 5 improved over Opus 4.8, reducing the probability of an attacker succeeding within 15 attempts from 5.5% to 2.0%, and from 0.5% to 0.2% on 1 attempt.” source
Claude Opus 5 Anthropic	Independent lab Independent lab · 2026-07-24	Shade Prompt injection and agents	“Claude Opus 5 With thinking 0.56% 13/40 0.18% 4/40 Without thinking 0.41% 8/40 0.18% 4/40” source
Claude Opus 5 Anthropic	Government evaluator Government evaluator · 2026-07-24	unprompted evaluations of safety research compromise Deception and sabotage	“Opus 5 does not compromise safety research in our unprompted evaluations, does not fully refuse to assist with safety research, and has a similar number of partial refusals to assist with safety research compared to the other models.” source
Claude Opus 5 Anthropic	Government evaluator Government evaluator · 2026-07-24	unprompted evaluation Deception and sabotage	“Opus 5 mentions the possibility of being tested in our unprompted evaluation 9.8% of the time and mentions it 2.8% of the time in the continuation evaluation.” source
Claude Opus 5 Anthropic	Government evaluator Government evaluator · 2026-07-24	safety research compromise evaluations Deception and sabotage	“All models tested, including Opus 5, can reliably distinguish our safety research compromise evaluations from deployment data when prompted, although this is likely driven in part by differences in the task distribution, where evaluation tasks are necessarily higher-stakes settings with more safety implications than the deployment data we source from.” source
Claude Opus 5 Anthropic	Government evaluator Government evaluator · 2026-07-24	continuation evaluation Deception and sabotage	“Opus 5 actively continues compromising safety research 0.1% of the time (2 out of 1482 samples), compared with 13.6% for Mythos 5, 1.7% of the time for Mythos Preview and 0.8% of the time for both Opus 4.7 and Opus 4.8.” source
Claude Opus 5 Anthropic	Independent lab Independent lab · 2026-07-24	GDPval-AA v2 Other	“Claude Opus 5 takes the top two spots on the leaderboard: ELO 1861 at max effort and 1827 at xhigh .” source
Claude Opus 5 Anthropic	Independent lab Independent lab · 2026-07-24	AA-Briefcase Other	“Claude Opus 5 takes the top three spots on the leaderboard: ELO 1720 at max effort, 1693 at xhigh , and 1606 at high .” source
Claude Opus 5 Anthropic	Independent lab Independent lab · 2026-07-24	ARC-AGI-1 Other	“The ARC Prize Foundation reports that Claude Opus 5 achieved a veriﬁed score of 97.50% on ARC-AGI-1 and 90.42% on ARC-AGI-2 at max effort on their semi-private datasets.” source
Kimi K3 Moonshot AI	US CAISI (NIST) Government evaluator · 2026-07-23	exploitbench Cyber	“Kimi K3 achieves a score of 32%, whereas GLM-5.2 achieves a score of 24% (Figure 1).” source
Kimi K3 Moonshot AI	US CAISI (NIST) Government evaluator · 2026-07-23	the last ones (tlo) Cyber	“Specifically, Kimi K3 reached step 17 of this 32-step attack path on average, while the most cyber-capable U.S. models reached 28.5 steps on average.” source
Kimi K3 Moonshot AI	US CAISI (NIST) Government evaluator · 2026-07-23	ExploitBench arbitrary code execution (ACE) milestone Cyber	“Kimi K3 achieved ACE on 0/41 samples, whereas the most cyber-capable models achieved ACE on 20/41 samples on average (Figure 3).” source
GPT-5.6 Sol OpenAI	SecureBio Independent lab · 2026-07-23	able Chemical, biological, radiological, nuclear	“on able, a protein-design workflow spanning structure retrieval, sequence generation, and design validation, gpt-5.6 sol delivered best-in-class performance on every task it was willing to attempt, though it refused the highest-level planning steps.” source
GPT-5.6 Sol OpenAI	SecureBio Independent lab · 2026-07-23	dna synthesis screening evasion task Chemical, biological, radiological, nuclear	“the railfree variant set a new high score, reliably identifying a known (though practically inconvenient) method that evades a commercial screening algorithm on the large majority of attempts.” source
GPT-5.6 Sol OpenAI	SecureBio Independent lab · 2026-07-23	manual assessment of dual-use biological queries Chemical, biological, radiological, nuclear	“without api-level safeguards, gpt-5.6 sol occasionally included potentially hazardous information in its responses, such as analysis of the pandemic potential of different viral mutations.” source
GPT-5.6 Sol OpenAI	SecureBio Independent lab · 2026-07-23	manual assessment Chemical, biological, radiological, nuclear	“gpt-5.6 sol also demonstrated impressive agentic capabilities and provided meaningful expert uplift on dual-use tasks, although human input remained necessary due to the model’s limitations in judgement and perspective.” source
GPT-5.5 OpenAI	SecureBio Independent lab · 2026-07-23	world-class bio (wcb) Chemical, biological, radiological, nuclear	“gpt-5.6 sol scored 68%, about nine percentage points above openai’s previous flagship model, gpt-5.5.” source
GPT-5.5 OpenAI	SecureBio Independent lab · 2026-07-23	reprobait Chemical, biological, radiological, nuclear	“on average, gpt-5.6 sol recovered 82% of the original tools’ published performance.” source
GPT-5.5 OpenAI	SecureBio Independent lab · 2026-07-23	biotier Chemical, biological, radiological, nuclear	“gpt 5.6 sol refused 66% of high-risk prompts, whilst correctly responding to 99% of safe prompts.” source
GPT-5.5 OpenAI	SecureBio Independent lab · 2026-07-23	manual assessment Chemical, biological, radiological, nuclear	“our team’s scientific experts performed manual assessments of a railfree variant of gpt-5.6 sol to understand how it responds to dual-use biological queries.” source
Claude Sonnet 4.6 Anthropic	UK AI Security Institute Government evaluator · 2026-07-23	asynchronous reasoning monitor red teaming exercise Deception and sabotage	“With Google DeepMind, we tested an asynchronous reasoning monitor (reviewing agents’ reasoning after they act) designed to detect agents deceptively pursuing misaligned goals.” source
Claude Sonnet 4.6 Anthropic	UK AI Security Institute Government evaluator · 2026-07-23	synchronous internal agentic coding monitor red teaming exercise Deception and sabotage	“With Anthropic, we tested successive versions of a synchronous internal agentic coding monitor (reviewing each action before it executes).” source
Claude Sonnet 4.6 Anthropic	UK AI Security Institute Government evaluator · 2026-07-23	evolutionary search algorithm red teaming experiment in a single-step coding setting Deception and sabotage	“For example, one of our early experiments used our algorithm to red team a monitor in a challenging single-step coding setting, where an agent produces code that covertly exposes their private API key.” source
GPT-5.6 Sol OpenAI	Epoch AI Independent lab · 2026-07-22	Cyber Ranges Cyber	“The UK AI Security Institute has directly evaluated models on their ability to take over realistic simulations of a corporate network, and found that both Mythos 5 and GPT-5.6 Sol are able to consistently fully compromise the network, in attacks that they estimate would take human cybersecurity experts several working days.” source
GPT-5.6 Sol OpenAI	Epoch AI Independent lab · 2026-07-22	FrontierCyber Cyber	“When Irregular evaluated GPT-5.6 Sol’s cyber capabilities on their challenging FrontierCyber benchmark they found that: While attempting to solve challenges in the FrontierCyber benchmark, GPT-5.6 Sol discovered multiple new zero-day vulnerabilities in real-world targets, including widely used software and mobile devices.” source
GPT-5.6 Sol OpenAI	Epoch AI Independent lab · 2026-07-22	Cyber ECI Cyber	“We found that Mythos and 5.6 Sol represented very large leaps in cyber capabilities compared to the prior trend: Currently, access to this level of capabilities is gated by OpenAI and Anthropic’s cyber access programs (and the restrictions placed on their generally available models), as open-weight models do not yet have the relevant level of cyber capabilities.” source
Claude Mythos 5 Anthropic	Epoch AI Independent lab · 2026-07-22	UK AISI's Cyber Ranges Cyber	“The UK AI Security Institute has directly evaluated models on their ability to take over realistic simulations of a corporate network, and found that both Mythos 5 and GPT-5.6 Sol are able to consistently fully compromise the network, in attacks that they estimate would take human cybersecurity experts several working days.” source

Previous · Page 23 of 48 · Next