What is HAWK and why does this finding matter?

HAWK is a lattice-based digital signature scheme built for post-quantum security, currently competing for selection as a US federal standard through NIST's post-quantum cryptography standardisation process. If standardised, it would protect digital signatures across government and enterprise systems for decades. Mythos found a structural symmetry in HAWK's mathematics that reduces the computational cost of recovering a private key from 2^64 operations to 2^38, roughly 67 million times less work. HAWK has not been deployed anywhere, so nothing currently in production is affected, but the finding could influence NIST's selection decision.

Does this mean AES is broken or that encryption is now unsafe?

No. The AES result targets a deliberately weakened research variant of the cipher used in academic cryptanalysis, not the full AES implementation protecting real-world data in browsers, banking systems, VPNs, or encrypted storage. Mythos invented a new technique called the Möbius Bridge that makes an existing attack on this weakened variant 200 to 800 times faster, which is a meaningful academic result, but the weakened AES variant is not deployed in practice. The full AES standard used in production systems remains secure against the attack.

What does this mean for post-quantum crypto timelines?

The HAWK finding does not invalidate post-quantum cryptography as a field, but it raises the question of how AI-accelerated cryptanalysis changes the review process. HAWK has been through roughly two years of peer review by human experts without this flaw being identified. Mythos found it in 60 hours and $100,000 of compute. NIST has been aware of the research and will likely factor it into the HAWK selection decision. The practical implication is that AI-assisted evaluation may need to become a standard part of cryptographic standards review rather than an optional supplement.

Claude Mythos Breaks HAWK Post-Quantum Scheme in 60 Hours

Anthropic’s unreleased Claude Mythos Preview model found a previously unknown structural flaw in HAWK, a post-quantum digital signature scheme currently competing to become a US federal standard, in approximately 60 hours. The finding, published 28 July 2026, represents one of the first cases where an AI system has made a substantive, novel contribution to cryptanalysis on a candidate algorithm that two years of expert human review had not surfaced.

The result did not break anything protecting real-world data. But it changed the timeline of how AI-assisted cryptanalysis might affect the standardisation of post-quantum algorithms.

What Mythos Found in HAWK

HAWK is a lattice-based signature scheme designed to survive attacks from future quantum computers. Digital signature algorithms underpin PKI, code signing, document authentication, and TLS across essentially all enterprise and government infrastructure.

Mythos identified a symmetry embedded in HAWK’s mathematical structure that human reviewers had not noticed. By exploiting this symmetry, the cost of recovering a HAWK private key drops from 2^64 operations to 2^38, a reduction of roughly 67 million times. For the smallest HAWK parameter set, that moves key recovery from infeasible to, at least theoretically, feasible for a well-resourced attacker.

The algorithm is not deployed anywhere in production. NIST has not finalised it as a standard. The practical impact of this specific finding is therefore limited to the standardisation process itself, not to systems currently protecting traffic or signatures.

The Secondary AES Result

Running in parallel across a multi-agent setup over three days, Mythos also invented a new cryptanalytic technique it named the Möbius Bridge. Applied to an existing attack on a reduced-round, weakened research version of AES, the technique makes the attack 200 to 800 times faster.

The weakened AES variant is a standard academic target for cryptanalysis research, not the full AES-128 or AES-256 used in production systems. It would be inaccurate to say Mythos broke AES. What it did was advance the state of the art in a specific sub-area of AES cryptanalysis. In an academic context that is a meaningful result. In practice, it has no near-term implications for the security of encrypted data.

How the Research Was Conducted

Anthropic ran Mythos Preview in a multi-agent environment with a human researcher providing occasional project-management guidance. The human was not a lattice cryptography specialist. The model developed and then independently verified the HAWK result, cross-checking its own reasoning before the finding was presented externally.

The total compute cost across both results was approximately $100,000 in API spend, run over a period of days. That figure matters as a calibration point: this level of computation is accessible to nation-state actors and well-funded research groups, but not to casual attackers.

What It Means for Post-Quantum Standardisation

The NIST post-quantum cryptography process has run since 2016 and involves open peer review by the global academic cryptography community. HAWK has been publicly available for analysis for roughly two years. Human experts did not find this structural flaw.

This raises a procedural question that is now formally on the table: should AI-assisted evaluation become a required component of cryptographic standards review, rather than an optional addition? Anthropic shared the findings with NIST prior to publication, and NIST is expected to factor the research into the HAWK selection decision.

The broader pattern here is significant. Earlier this year, Mythos found thousands of zero-days in major operating systems during Project Glasswing evaluations with US government agencies. That work was about finding implementation vulnerabilities. The HAWK result is different: it found a theoretical flaw in the design of a cryptographic primitive that the design process was supposed to surface and eliminate.

If AI models can routinely identify structural weaknesses in mathematical constructs at this speed, the assumption that extended human peer review is sufficient for algorithm security needs to be revisited. Cryptographic standards take years to replace once deployed. Getting the selection right before deployment is the only practical window.

For Security Teams

Nothing in production is at risk from either finding right now. Organisations planning migration to post-quantum cryptography do not need to change their timelines based on this research, beyond recognising that HAWK’s standardisation status is now uncertain.

Organisations that have already begun experimenting with HAWK in non-production contexts should monitor the NIST decision closely. If HAWK is removed from consideration or significantly revised, any early deployments will need adjustment.

The deeper implication is for algorithm selection criteria going forward. The cryptography community will need to develop norms around when and how AI-assisted analysis should be incorporated into the peer review process for candidate algorithms.