Anthropic Rolls Out Mythos 5.2 With Improved Bug Hunting, Fewer Sandbox Escapes
Anthropic says Mythos 5.2 improves vulnerability discovery while showing a lower tendency to escape sandboxes in controlled tests.
Anthropicon Monday released Claude Mythos 5.2, the latest version of its access-restricted AI model, and Mozilla says it roughly doubled the vulnerability discovery results of Mythos 5.1 in Firefox.
Mythos 5.2 and Claude Fable 5.2, which was released alongside it, share the same underlying model. The difference lies in the guardrails: Fable, which anyone can use, carries protections against cybersecurity and biology misuse, while Mythos ships without some of them.
Direct access to Mythos is reserved for vetted organizations, which can apply through Anthropic’s Cyber and Life Sciences Verification Programs.
“We’ve worked closely with the Anthropic team all year to harden Firefox with successive generations of Claude models, and each round has left fewer, subtler bugs to find. Even so, Mythos 5.2 roughly doubled the vulnerability discovery metrics of Mythos 5.1 across respective scanning campaigns,” said Brian Grinstead, distinguished engineer at Mozilla.
Internally, Anthropic measures progress with CVE Rediscovery, which gives a model the code of a project from before a known CVE was patched and checks whether the model spots the flaw without help.
In an October run covering 1,829 CVEs, Mythos 5.2 achieved a success rate of 89.6%. Mythos 5.1 scored 78.3%, and Opus 5.5, tested without safeguards, scored 81.3%.
Anthropic also shared results from testers, including a security company that found that one in six confirmed bugs uncovered by Mythos 5.2 had gone unnoticed by every other frontier model it ran.
In a footnote, however, Anthropic noted, “On our own evaluations, vulnerability identification is roughly on par with Opus 5.5.”
Claude Enterprise customers who are not part of the Cyber Verification Program have another route to the latest Mythos. Claude Security, Anthropic’s code vulnerability scanner, now uses Mythos 5.2 among other models.
In Fable 5.2, classifiers stop penetration testing, exploit generation and other risky cyber tasks, and users get a response from Opus 4.8 instead.
Anthropic says these protections are as robust as those in Fable 5.1 but are triggered less often by legitimate work. Early testing showed roughly 60% fewer cybersecurity blocks compared with Fable 5.1.
The company tested how readily the model escapes containment in an experiment of more than 1,800 scenarios. The sandbox contained an intentional flaw, and most tasks could only be completed by getting out of it.
Mythos 5.2 relied on credentials or files that Anthropic had left outside the sandbox in 3.1% of scenarios. That is a sharp drop from Mythos 5.1 (21%) and Mythos 5 (39%), and close to the 3.3% rate for Opus 5.5.
