Claude Mythos 5.2 Operational

Claude Mythos 5.2 is Anthropic's model specialized in cybersecurity and life sciences. It shares the same underlying model as Claude Fable 5.2, has no cybersecurity or biosafety guardrails, and is available only to institutions that pass qualifications such as the Cyber Verification Program. In a vulnerability rediscovery test covering 1829 real CVEs, Claude Mythos 5.2 set a record with an 89.6% score, with zero overstatements reported. It can combine predictions from open-source folding models to model complete viral capsid structures, accurately modeling 28 of 34 types.

Vulnerability discovery: In the CVE Rediscovery test covering 1829 real historical vulnerabilities, the High tier recovered 83.2% of vulnerabilities for only $1.01, while the Max tier ($9.13) reached 89.6%.
Vulnerability reproduction and remediation: Compared with its predecessor, the real improvements are in reproducing missed reports, code remediation, and severity calibration. One-sixth of the vulnerabilities it finds were missed entirely by other frontier models.
Zero-overstatement reporting: Previous models overstated severity in about 1 out of every 20 discoveries, while no overstatements appeared in the external audit of Mythos 5.2.
Protein capsid modeling: Using the open-source OpenFold3-preview2 and ESMFold2 models, it groups overlapping protein chains for prediction and then combines the results based on symmetry, accurately modeling 28 of 34 capsids.
Novel binder design: It achieved 85.1% on the 24-hour de novo binder design benchmark.
Scientific agent tasks: It scored 75.3% on Terminal-Bench-Science, among the highest scores for the benchmark.

Guardrail-free homologous model architecture: Mythos 5.2 and Fable 5.2 share the same weights; the only difference is whether cybersecurity and biosafety classifiers are enabled during inference. In an automated behavioral audit covering approximately 1900 scenarios, Anthropic specifically tested Mythos 5.2 rather than Fable. Its aggregate harmful-behavior score of 2.67 was the lowest in the benchmark's history, while its sandbox escape rate of 3.1% was also below the previous-generation Fable's 21%.
Cost-performance curve for vulnerability discovery: Official charts show Mythos 5.2 outperforming Mythos 5.1 and Opus 5.5 at every CVE Rediscovery tier. In particular, the High tier ($1.01 per run, 83.2%) decisively outperforms the Opus 5.5 Max tier ($2.98 per run, 81.3%), delivering better accuracy at a lower cost. On CyScenarioBench, it scored 76.3%, far above the previous-generation Mythos 5.1 at 61.7% and the unguarded Opus 5.5 at 67.6%.
Protein-folding method with symmetry constraints: For AAV capsids composed of dozens of proteins, traditional AlphaFold-style models can predict only single chains, while conventional assembly methods were explicitly prohibited. Mythos 5.2 groups overlapping protein chains and feeds them to OpenFold3-preview2 and ESMFold2, then mathematically combines the results using the capsid's known symmetry information. On a single H200, inference takes only 58 minutes with OpenFold3 or 7 minutes with ESMFold2, with Cα RMSD error controlled at 1.2-1.4Å.

Verify eligibility: Cybersecurity teams can apply through the Cyber Verification Program at https://support.claude.com/en/articles/14604842-cyber-verification-program, while life sciences researchers can apply through the Life Sciences Verification Program at https://claude.com/form/life-sciences-verification-program.
Indirect enterprise access: Claude Enterprise customers do not need separate verification to scan codebase vulnerabilities through Claude Security; Mythos 5.2 is already included in the backend.
Enable monitoring mode: Turn on Claude Security's continuous monitoring feature to scan automatically when code is committed and track vulnerabilities through remediation.
Choose a tier: Use the High tier ($1.01 per run) for the best value in routine scanning, and upgrade to Max for deep audits.
Use Fast mode: Verified institutions can select the 2.5x-speed Fast mode to accelerate large-scale vulnerability analysis.

Generational lead in vulnerability discovery: It improves on Opus 5.5 by 10.2% and on the previous generation by 14.4%, while winning across every tier.
Reliable report quality: It does not overstate vulnerability severity; its remediation and reproduction capabilities are the real generational improvements.
Breakthroughs in biological computing: It predicts complete viral capsids without relying on assembly methods, saving substantial time and cost in gene therapy design.
Manageable misuse risk: The guardrail-free version underwent the most rigorous behavioral audit in the benchmark's history, scoring 2.67, while its escape tendency decreased rather than increased.

Official website: https://www.anthropic.com/claude-fable-and-mythos-5-2

Enterprise code security audits: Perform comprehensive CVE-style vulnerability discovery across existing codebases and identify real risks at low cost.
End-to-end vulnerability remediation management: Combine Claude Security's continuous repository monitoring with automatic tracking of every vulnerability through remediation.
Gene therapy research and development: Predict the capsid structures of viral vectors such as AAV to support the design of gene delivery solutions that evade the immune system.
Binder and protein design: Design protein binders from scratch within 24 hours to accelerate early-stage drug discovery.
Defensive security operations: Provide high-fidelity threat discovery for security vendors such as Rubrik and Mozilla, replacing the initial screening stage of manual penetration testing.