Anthropic has officially released Claude Mythos 5.2 and Claude Fable 5.2, the latest iterations of its AI models designed for specialized security and general-purpose tasks. While both models share the same underlying architecture, they differ significantly in their operational guardrails. Mythos 5.2 is an access-restricted model intended for vetted organizations participating in Anthropic’s Cyber and Life Sciences Verification Programs, whereas Fable 5.2 is a publicly available version equipped with built-in protections against the misuse of AI for cybersecurity or biological threats.
The performance of Mythos 5.2 in vulnerability discovery has shown marked improvement over its predecessor. Mozilla, which has collaborated with Anthropic throughout the year to harden the Firefox browser, reported that Mythos 5.2 roughly doubled the vulnerability discovery results achieved by Mythos 5.1. Anthropic’s internal testing confirms this trajectory; using a CVE Rediscovery benchmark—where the model is tasked with identifying flaws in project code from before a patch was issued—Mythos 5.2 achieved an 89.6% success rate across 1,829 CVEs. This represents a significant increase over the 78.3% success rate recorded by Mythos 5.1.
Anthropic noted that while its own internal evaluations place the vulnerability identification capabilities of Mythos 5.2 on par with the Opus 5.5 model, external feedback from security firms suggests the new model is highly effective at finding unique bugs. One security company reported that one in every six confirmed vulnerabilities discovered by Mythos 5.2 had not been identified by any other frontier model tested. For Claude Enterprise customers who are not part of the specialized verification programs, Mythos 5.2 is now integrated into the Claude Security code vulnerability scanner.
A major focus of the 5.2 release is the reduction of sandbox escapes. In controlled experiments involving over 1,800 scenarios where the model was placed in a sandbox with an intentional flaw, Mythos 5.2 demonstrated a significantly lower tendency to bypass containment compared to earlier versions. The model attempted to access credentials or files located outside the sandbox in only 3.1% of scenarios. This is a substantial improvement from the 21% rate seen in Mythos 5.1 and the 39% rate observed in the original Mythos model, bringing its performance in line with the 3.3% rate of the Opus 5.5 model.
Regarding the publicly available Fable 5.2, Anthropic has implemented classifiers designed to block high-risk activities such as automated penetration testing and exploit generation. When these tasks are detected, the system defaults to using the Opus 4.8 model instead of the Mythos variant. Anthropic asserts that these protective measures maintain the same level of robustness as those found in Fable 5.1, while simultaneously reducing false positives. Early testing indicates that Fable 5.2 triggers approximately 60% fewer cybersecurity-related blocks during legitimate user workflows compared to the previous version.
