Superior Performance in Software Vulnerability Discovery
Anthropic has announced the launch of the new version, Claude Mythos 5.2, a restricted-access AI model dedicated to advanced security tasks. Practical tests have shown a radical improvement in the model's ability to discover vulnerabilities; Mozilla reported that the new version nearly doubled vulnerability discovery results in the Firefox browser compared to the previous version (Mythos 5.1).
In an internal company-approved test known as CVE Rediscovery—which measures the model's ability to identify previously known vulnerabilities in software code before they are patched—Mythos 5.2 achieved a success rate of 89.6% out of 1,829 tested vulnerabilities (CVEs). Interestingly, this performance outperformed the Opus 5.5 model, which was tested without security restrictions, as the latter recorded only 81.3%.
Enhancing Security and Reducing Sandbox Escape Risks
Alongside improving offensive capabilities (for testing purposes), Anthropic focused on enhancing security and control aspects. Experiments showed a sharp decrease in the "Sandbox Escape" rate; the model did not escape the testing environment in only 3.1% of scenarios, compared to 21% in version 5.1 and 39% in version 5. This development is crucial to ensuring that penetration testing and security research remain within a controlled and secure scope.
In parallel, the company launched the Claude Fable 5.2 version, available for general use, which shares the same base model as Mythos but possesses stricter guardrails to prevent misuse in the fields of cybersecurity or biology. Anthropic has succeeded in reducing "false alarms" in this version, recording a 60% decrease in the blocking of legitimate security tasks compared to the previous version.
What Changes in Practice?
This update represents a significant shift in how artificial intelligence is used in the Secure Software Development Life Cycle (DevSecOps). For organizations certified in Anthropic's security and bioscience verification programs, Mythos 5.2 provides a more precise and reliable tool for vulnerability hunting without the risk of the model losing control within testing environments.
As for developers and regular users via the Fable version, the update means a smoother experience with reduced unnecessary interventions from protection systems, allowing technical research to be conducted without legitimate activities being classified as security threats. Anthropic has also already begun integrating Mythos 5.2 into the Claude Security tool, a dedicated software vulnerability scanner.