Anthropic has introduced the latest iteration of its AI model, Mythos 5.2, with significant improvements in identifying vulnerabilities. Released on Monday, this version has nearly doubled the bug discovery capabilities compared to its predecessor, Mythos 5.1, specifically in Mozilla’s Firefox, according to company reports.
Enhanced Vulnerability Detection
Mythos 5.2, along with Claude Fable 5.2, is built on the same foundational model. However, Claude Fable is designed with additional safeguards to prevent cybersecurity and biological misuse, whereas Mythos 5.2 is available without these controls to vetted organizations through Anthropic’s Cyber and Life Sciences Verification Programs.
Mozilla’s collaboration with Anthropic has bolstered Firefox’s defenses, as each successive Claude model generation has progressively minimized the presence of bugs. Brian Grinstead, a distinguished engineer at Mozilla, noted that Mythos 5.2 has significantly outperformed Mythos 5.1 in vulnerability detection metrics during recent scanning campaigns.
Performance Metrics and Evaluations
Anthropic evaluates its models using the CVE Rediscovery methodology, which tests a model’s ability to identify known vulnerabilities in code pre-patch. In a comprehensive test involving 1,829 CVEs, Mythos 5.2 achieved an impressive 89.6% success rate, surpassing the 78.3% of Mythos 5.1 and the 81.3% of Opus 5.5, which was evaluated without its usual safeguards.
Moreover, external testers, including a security firm, reported that Mythos 5.2 uncovered one in six confirmed bugs that had not been detected by any other advanced model. However, Anthropic also mentioned that on their internal assessments, the vulnerability identification capabilities of Mythos 5.2 are comparable to Opus 5.5.
Improved Security Measures
For those not part of the Cyber Verification Program, Claude Enterprise clients can access the newest Mythos version via Claude Security, which incorporates Mythos 5.2 among other models for scanning code vulnerabilities. In Claude Fable 5.2, additional measures prevent risky activities such as penetration testing and exploit generation, defaulting to responses from Opus 4.8 instead.
Anthropic’s internal testing showed a reduction in cybersecurity blocks by approximately 60% compared to Fable 5.1, while maintaining robust protection levels. The company also conducted sandbox escape tests across more than 1,800 scenarios, revealing that Mythos 5.2 relied on external credentials or files in just 3.1% of cases, a significant drop from previous versions.
These advancements underscore Anthropic’s commitment to refining AI models for enhanced cybersecurity, demonstrating a promising trajectory for future developments in AI-driven vulnerability detection.
