Anthropic's Mythos Model Discovers Security Vulnerabilities Faster Than Teams Can Patch

Anthropic's Claude Mythos model, deployed under Project Glasswing, has broken a NIST post-quantum cryptography candidate: Mythos found a flaw in the HAWK digital signature algorithm that caused its developer to withdraw it from NIST's standardization process after HAWK had passed two evaluation rounds [5]. Separately, the program has reported 23,000 potential vulnerabilities across 1,000 open-source software projects [2], and Microsoft has held emergency engineering meetings as Mythos discovers flaws faster than teams can patch them [1]. Cryptographers are now debating whether AI-assisted cryptanalysis during the PQC transition period is net positive or a dual-use liability [6][1].

The HAWK withdrawal shows that AI-assisted cryptanalysis can produce concrete consequences for security standards — a NIST PQC candidate that survived two human-led evaluation rounds was broken and removed from contention before widespread deployment. The dual-use concern runs through every part of this: the same capability that caught a flaw in a candidate standard could be turned toward breaking standards already in use.

Anthropic's Claude Mythos model has been deployed under Project Glasswing to find software vulnerabilities in critical systems at a rate that has exceeded human patching capacity. Microsoft was among the first partners; the company convened emergency engineering meetings to handle vulnerabilities Mythos was uncovering faster than its teams could address [1]. Anthropic frames the program as a race to find and fix flaws before adversarial governments — China is specifically named — deploy equivalent AI tools [1]. By mid-2026, Mythos had reportedly identified over 23,000 potential vulnerabilities across 1,000 open-source software projects [2], growth from the 10,000 reported in May [3].

In parallel, Anthropic researchers ran Mythos through a focused 60-hour cryptographic research session at approximately $100,000 in API costs, targeting mathematical weaknesses in the HAWK post-quantum signature standard and a reduced-round version of AES-128 [4]. A distinctive feature of the session was that human interventions were primarily motivational rather than technical: researchers repeatedly prompted Mythos not to conclude problems were unsolvable, urging it to find something publishable [4]. The model produced genuine mathematical analysis, finding weaknesses that had not surfaced during years of expert evaluation.

The HAWK findings proved more consequential than initially characterized. HAWK had passed two rounds of NIST's post-quantum cryptography evaluation before Mythos identified a flaw severe enough to render the algorithm broken [5]. Following Anthropic's public disclosure, HAWK's developer withdrew the algorithm from NIST's standardization process entirely [5]. The result is concrete: a candidate standard designed to secure communications against future quantum computers was removed from consideration by AI-assisted cryptanalysis.

The results have drawn reactions from the security and cryptography communities. Matthew Green, quoted by Simon Willison, argues the timing is close to ideal — the PQC transition period, before any of these algorithms are widely deployed, is exactly when catching weaknesses is most valuable, and AI cryptanalysis arriving now is net positive for defenders [6]. That view sits alongside the investigative reporting's dual-use concern: the same capability that helped remove a flawed standard also tells adversarial programs what to build [1]. Internally, at least one Microsoft engineer questioned whether Mythos lived up to Anthropic's capability claims [1], a gap between the company's public cooperation with Project Glasswing and private assessments of the model's performance that remains unresolved.