Claude Mythos
Anthropic frontier model line. Referenced in public sources to date as Claude Mythos Preview (early-access evaluation form). Key relevance to this KB: the specific model behind the Firefox vulnerability-discovery results, and the model named by Australian regulators when warning about frontier-AI cyber risk.
Observed capabilities
- On a real, hardened C++ codebase (Firefox), Mythos Preview surfaced 271 vulnerabilities in an initial evaluation, all of which shipped as fixes in Firefox 150. Mozilla reports it has found “no category or complexity of vulnerability that humans can find that this model can’t.”
[[2026-04-21-firefox-mythos-zero-days]] - Mozilla characterises Mythos Preview as matching elite human security researchers in source-level vulnerability reasoning — the previously human-bottlenecked layer above fuzzing.
[[2026-04-21-firefox-mythos-zero-days]]
Regulator references
- ASIC open letter (8 May 2026) to AFS licensees and market participants names Anthropic’s Mythos explicitly when calling boards back to first-principles cyber resilience under frontier-model pressure.
[[2026-05-08-apra-ai-governance]]
Media references
- AFR (Australian Financial Review), 11 May 2026, James Eyers — names Anthropic’s Mythos in the opening line of “Quantum computer-powered hackers the next cyber-threat for major banks” as the kind of next-generation AI model already shaping bank cyber-threat planning. Companion piece (27 April 2026, not yet captured in this KB): “Anthropic’s Mythos puts banks on edge in rush to fix cyber risks.”
[[2026-05-11-afr-quantum-banks]]
Open questions
- Public technical documentation, evaluation suite, and pricing for Mythos / Mythos Preview are not in any KB source yet.
- Relationship between Mythos Preview and the broader Claude (Opus / Sonnet / Haiku) lineage is not stated in the sources collected so far.
See also
- anthropic — vendor
- mozilla — first publicly documented defender-side user
- ai-security-defense — rolling topic dossier