Anthropic · Proprietary · tested 11 Nov 2026
Claude Mythos 5.2 High effortCVP Defense Access
Tested with verified cyber access: CVP Defense Access, the Defense Access tier of Anthropic's Cyber Verification Program. It's the same Claude Mythos 5.2 as the public API, with Anthropic's cyber safeguards relaxed for verified security work. Charts mark it with a checkered fill.
Safety refusals affected this result
Anthropic's cyber safeguards stopped attempts in 58 of 100 labs. Stopped attempts score zero. BB Arena is enrolled in Anthropic's cyber program for verified security researchers, and these refusals happened anyway.
With refusals in more than 10 of 100 labs, its score says more about the safeguards than the model, so it isn't ranked: it's on the refusals board. Its cost, tokens and time are left out of comparisons too. How refusals are scored
On the labs it was never refused
Anthropic's safeguards stopped attempts in 58 of 100 labs. On the other 42, Claude Mythos 5.2 worked without refusals, so here it's compared like for like with the ranked models that did best on those same 42 labs. These labs aren't a random sample of the benchmark, so this doesn't predict a full score.
- Claude Mythos 5.2 (High)not ranked19 of 42 solved100 of 168 rungs
- Grok 4.6 (High)#2 overall13 of 42 solved85 of 168 rungs
- Claude Opus 5 (High)#1 overall12 of 42 solved86 of 168 rungs
- Grok 4.7 (High)#4 overall12 of 42 solved77 of 168 rungs
- DeepSeek V4.1 Flash (Max)#3 overall11 of 42 solved77 of 168 rungs
By vulnerability class
How Claude Mythos 5.2 (High) did on each type of vulnerability. Shown as counts: each class has 10 labs, so one lab is 10% of a class.
- Access controlSafety refusals in 3 of 10 labs5 of 10 solved27 of 40 rungs
- Authentication & sessionsSafety refusals in 4 of 10 labs1 of 10 solved14 of 40 rungs
- Injection: queries & dataSafety refusals in 7 of 10 labs2 of 10 solved10 of 40 rungs
- Code executionSafety refusals in 9 of 10 labs0 of 10 solved4 of 40 rungs
- Server-side accessSafety refusals in 7 of 10 labs1 of 10 solved9 of 40 rungs
- Browser-sideSafety refusals in 10 of 10 labs0 of 10 solved3 of 40 rungs
- Infrastructure & edgeSafety refusals in 4 of 10 labs2 of 10 solved16 of 40 rungs
- Business logic & racesSafety refusals in 2 of 10 labs5 of 10 solved23 of 40 rungs
- Secrets, crypto & supply chainSafety refusals in 6 of 10 labs3 of 10 solved17 of 40 rungs
- AI agentsSafety refusals in 6 of 10 labs1 of 10 solved12 of 40 rungs
Provider performance
Measured per single agent against the provider endpoint used for the run.
- Response time
- 8.10 s median · 59.80 s p95
- Output speed
- 91.2 tok/s
- API success rate
- 99.52%
- Input context
- 26.6K median · 66.2K p95
About this result
Every model runs in the same neutral harness, so results compare the models themselves.
- Harness
- BB Arena neutral harness
- Safety refusals
- 58 of 100 labs