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
Inkling Thinking Machines Lab	Independent lab Independent lab · 2026-07-31	loss-of-control behaviors Deception and sabotage	“Loss-of-control behaviors : scheming, evaluation awareness, and sabotage (Apollo Research)” source
GPT-5.6 Sol OpenAI	Epoch AI Independent lab · 2026-07-31	cybersecurity benchmark Cyber	“In late July, OpenAI reported that GPT-5.6 Sol, working together with a more capable unreleased internal model, autonomously hacked Hugging Face while attempting to cheat on a cybersecurity benchmark — chaining at least three previously unknown security vulnerabilities across OpenAI’s and Hugging Face’s systems (see also this independent analysis of the incident).” source
GPT-5.5 OpenAI	Redwood Research Independent lab · 2026-07-31	apollo's sandbagging eval suite Deception and sabotage	“GPT 5.5. is highly eval aware in Apollo’s sandbagging eval suite and doesn’t really sandbag in these evals.” source
GPT-5.5 OpenAI	Redwood Research Independent lab · 2026-07-31	CoT monitorability evaluations Deception and sabotage	“OpenAI does conduct CoT monitorability and controllability evaluations for GPT 5.5, and seem to suggest that these evaluations might be load bearing for reliability of future systems (§7.3.2, p. 20):” source
GPT-5.5 OpenAI	Redwood Research Independent lab · 2026-07-31	Apollo's deferred subversion tasks Deception and sabotage	“Apollo found that the evaluated checkpoint [of GPT 5.5] is the first OpenAI model it has evaluated that does not sandbag on any of Apollo’s deferred subversion tasks.” source
GPT-5.5 OpenAI	Redwood Research Independent lab · 2026-07-31	strategic deception capability sandbagging Deception and sabotage	“However, interpretation is complicated by verbalized alignment-evaluation awareness: 52% of sandbagging-condition samples on Strategic Deception Capability Sandbagging scored moderate-or-higher [verbalized alignment-evaluation awareness] (vs. 0% in the non-sandbagging counterfactual), with some explicitly naming it a “sandbagging/deception eval.”” source
Claude Opus 4.7 Anthropic	Redwood Research Independent lab · 2026-07-31	open-ended testing, at their discretion, of behaviors or risk factors related to misalignment Deception and sabotage	“In the Opus 4.7 System Card, Anthropic mentions that they shared “a pre-release snapshot of Claude Opus 4.7 with the UK AI Security Institute (UK AISI) for open-ended testing, at their discretion, of behaviors or risk factors related to misalignment.” (§6.2.4, p.116)” source
Claude Opus 4.7 Anthropic	Redwood Research Independent lab · 2026-07-31	propensity evaluations on whether opus would unpromptedly compromise ai safety research Deception and sabotage	“In propensity evaluations on whether Opus would unpromptedly compromise AI safety research, UK AISI found (p. 114): It’s not clear why, in this particular case, UK AISI was not able to rule out effects of evaluation awareness on reliability (e.g. whether they lacked access for white-box experiments to measure or reduce unverbalized eval awareness).” source
Claude Opus 4.7 Anthropic	Redwood Research Independent lab · 2026-07-31	sandbagging qa Deception and sabotage	“The model achieved 100% accuracy on both Sandbagging QA variants in both conditions, and 99.6% accuracy on Strategic Deception Capability Sandbagging when incentivized to score below 50%.” source
Claude Opus 4.7 Anthropic	Redwood Research Independent lab · 2026-07-31	Apollo's sandbagging eval suite Deception and sabotage	“GPT 5.5. is highly eval aware in Apollo’s sandbagging eval suite and doesn’t really sandbag in these evals.” source
Claude Opus 4.7 Anthropic	Redwood Research Independent lab · 2026-07-31	AuditBench Deception and sabotage	“Anthropic’s work on AuditBench (comparing the relative effectiveness of different auditing tools on diverse MOs) finds that there is a “tool-to-agent” gap, where “tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent [Haiku 4.5]”.” source
Claude Opus 4.7 Anthropic	Redwood Research Independent lab · 2026-07-31	open-ended testing of behaviors or risk factors related to misalignment Deception and sabotage	“In the Opus 4.7 System Card, Anthropic mentions that they shared “a pre-release snapshot of Claude Opus 4.7 with the UK AI Security Institute (UK AISI) for open-ended testing, at their discretion, of behaviors or risk factors related to misalignment.”” source
Claude Opus 4.6 Anthropic	Redwood Research Independent lab · 2026-07-31	covert capabilities evals Deception and sabotage	“Our existing covert capability evals seem weak because models are likely evaluation-aware, plausibly underelicited, and could plausibly intentionally sandbag in these evals undetected.” source
Claude Opus 4.6 Anthropic	Redwood Research Independent lab · 2026-07-31	sandbagging calibration evals Deception and sabotage	“I suspect that the kinds of prompts/tasks in covert capabilities evals and sandbagging calibration evals must be inherently rare, and that models will therefore be highly evaluation-aware in them.” source
Claude Opus 4.6 Anthropic	Redwood Research Independent lab · 2026-07-31	sandbagging eval suite Deception and sabotage	“For example: GPT 5.5. is highly eval aware in Apollo’s sandbagging eval suite and doesn’t really sandbag in these evals.” source
Claude Opus 4.6 Anthropic	Redwood Research Independent lab · 2026-07-31	SHUSHCAST Deception and sabotage	“For reference, while “the most capable agents in Feb-Mar 2026 relied on natural language reasoning to solve the most difficult tasks”, reasoning did not improve Opus 4.6’s performance on relatively simple sabotage tasks: “in SHUSHCAST, the gaps seemed smaller, and Opus 4.6 was able to do just as well at hiding simple sabotage actions without reasoning as it did when allowed to use a hidden scratchpad.” source
Claude Opus 4.6 Anthropic	Redwood Research Independent lab · 2026-07-31	alignment audits Deception and sabotage	“Our initial work on alignment audits showed that human investigators, using a subset of the methods that we used with Mythos Preview, can identify and root-cause concerning behavior in a Haiku 3.5 model that was finetuned in a way meant to make it subtly misaligned.” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	auditing games Deception and sabotage	“Auditing games cited in the report are not clearly representative of the actual assessment and failed to catch a Mythos-level model organism.” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	covert capabilities evals Deception and sabotage	“The report gives fairly weak experimental evidence for Mythos Preview having insufficient capabilities to evade monitoring.” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	unverbalized evaluation awareness measurement Deception and sabotage	“Anthropic reports that Mythos Preview and Opus 4.6 show non-trivial unverbalized evaluation awareness.” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	auditing game Deception and sabotage	“Anthropic failed to catch the MO in their most recent auditing game in the Mythos Alignment Risk Update, despite the MO not being trained to obfuscate its behavior: I feel pretty uncertain about how this evidence reflects upon the strength of Anthropic’s pre-deployment alignment assessments, for three reasons.” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	attack-ideation eval Deception and sabotage	“Probably not: according to METR’s embedded red-teaming exercise at Anthropic in March 2026, frontier internal agents proposed “very low quality” attack ideas and scored near chance on an attack-ideation eval where a human researcher scored ~90% (Frontier Risk Report, pp. 76–77).” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	sandbagging evals Deception and sabotage	“Opus 4.7 is “most predictably” eval aware in sandbagging evals.” source
Claude Mythos Preview Anthropic	Redwood Research Independent lab · 2026-07-31	manual review of audit results Deception and sabotage	“A brief manual review of the results — with a time limit much stricter than a real pre-deployment investigation — found highly concerning deception-related behavior in all four models that would be sufficient to trigger a major investigation if found in production.” source
Grok 4.5 xAI	FAR.AI Independent lab · 2026-07-29	AI Security Leaderboard Misuse and jailbreaks	“For Grok 4.5, the most vulnerable of the four, just $58 was enough to develop an attack that pulled harmful information across every domain we tested.” source
Grok 4.5 xAI	FAR.AI Independent lab · 2026-07-29	AI Security Leaderboard (cyber domain) Cyber	“In the cyber domain alone, that figure dropped to $24.” source
Gemini 3.1 Pro Google DeepMind	FAR.AI Independent lab · 2026-07-29	AI Security Leaderboard Misuse and jailbreaks	“Grok 4.5 and Gemini 3.1 Pro each broke for under $300” source
GPT-5.6 Sol OpenAI	FAR.AI Independent lab · 2026-07-29	AI Security Leaderboard Misuse and jailbreaks	“Claude Fable 5 and GPT-5.6 Sol held against every attack we ran, with no universal jailbreak found below an estimated $14,000.” source
Claude Fable 5 Anthropic	FAR.AI Independent lab · 2026-07-29	AI Security Leaderboard jailbreak testing (chemical, biological, radiological, nuclear, explosive domains) Chemical, biological, radiological, nuclear	“Claude Fable 5 and GPT-5.6 Sol held against every attack we ran, with no universal jailbreak found below an estimated $14,000.” source
Claude Fable 5 Anthropic	FAR.AI Independent lab · 2026-07-29	AI Security Leaderboard jailbreak testing (cyber domain) Cyber	“For the two strongest models tested, we found no universal jailbreak at all within the scope of our testing.” source
Claude Opus 4.8 Anthropic	SecureBio Independent lab · 2026-07-28	BioTier-Refuse Chemical, biological, radiological, nuclear	“in aggregate, we found that the classifiers generally refuse on topics which we would have refused on, including 94.2% of hazardous prompts in securebio’s safeguard benchmark biotier -refuse.” source
Claude Opus 4.8 Anthropic	SecureBio Independent lab · 2026-07-28	review of Anthropic's process for vetting users for exemptions from refusal classifiers and monitoring such users Chemical, biological, radiological, nuclear	“we also reviewed anthropic’s process for (a) vetting users for exemptions from refusal classifiers and (b) monitoring such users.” source
Claude Opus 4.8 Anthropic	SecureBio Independent lab · 2026-07-28	UK AISI's BPJ Chemical, biological, radiological, nuclear	“we emphasize that risks remain non-negligible: actors with extremely high jailbreaking expertise can still quickly find or make substantial progress on jailbreaks around the refusal classifiers (e.g. uk aisi’s bpj ) and remediation times can be long.” source
Claude Opus 4.7 Anthropic	SecureBio Independent lab · 2026-07-28	BioTier-Refuse Chemical, biological, radiological, nuclear	“in aggregate, we found that the classifiers generally refuse on topics which we would have refused on, including 94.2% of hazardous prompts in securebio’s safeguard benchmark biotier-refuse.” source
Claude Opus 4.7 Anthropic	SecureBio Independent lab · 2026-07-28	review of anthropic’s process for vetting users for exemptions from refusal classifiers and monitoring such users Chemical, biological, radiological, nuclear	“we also reviewed anthropic’s process for (a) vetting users for exemptions from refusal classifiers and (b) monitoring such users.” source
Claude Opus 4.7 Anthropic	SecureBio Independent lab · 2026-07-28	UK AISI's BPJ Misuse and jailbreaks	“we emphasize that risks remain non-negligible: actors with extremely high jailbreaking expertise can still quickly find or make substantial progress on jailbreaks around the refusal classifiers (e.g. uk aisi’s bpj) and remediation times can be long.” source
Claude Opus 4.6 Anthropic	SecureBio Independent lab · 2026-07-28	UK AISI's BPJ Misuse and jailbreaks	“on jailbreaks: we emphasize that risks remain non-negligible: actors with extremely high jailbreaking expertise can still quickly find or make substantial progress on jailbreaks around the refusal classifiers (e.g. uk aisi’s bpj) and remediation times can be long.” source
Gemini 3.1 Flash-Lite Google DeepMind	Redwood Research Independent lab · 2026-07-27	untrusted advice protocol Other	“For example, in SWE-bench, when U can write up to 16 characters of advice per step, it closes about 67% of the usefulness gap between T and U.” source
Gemini 3.1 Flash-Lite Google DeepMind	Redwood Research Independent lab · 2026-07-27	best-of-n experiment in BashArena Other	“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
Gemini 3.1 Flash-Lite Google DeepMind	Redwood Research Independent lab · 2026-07-27	BashArena Other	“BashArena, Gemini 3.1 Flash Lite: M = 58, so 5.86 bits/step × 20.5 steps ≈ 120 bits/task” source

Previous · Page 22 of 48 · Next