Claude Found a Weakness in Post-Quantum Signatures. No Quantum Computer Was Needed.

HackWednesday AI Security Desk3 min read

AI in SecurityAI-assistedOwner approved for publication5 linked sources

Sources checked . Incident guidance can change; follow the linked vendor advisories.

Anthropic’s Claude helped find a weakness in the HAWK signature candidate. What HAWK’s withdrawal means, what wasn’t broken, and how to plan a migration.

A cracked bronze signature seal on a research workbench.
AI-generated editorial illustration. The seal represents a weakness in a proposed signature scheme, not a production breach.
Editorial note: The site owner approved publication of this AI-assisted article without a separate human editorial review.
In this article (3 sections)

On July 28, 2026, Anthropic reported that researchers working with Claude Mythos Preview found a stronger way to attack HAWK, a proposed post-quantum digital signature scheme. The work used conventional computing, not a quantum computer. It was a research result, not a report of stolen keys or compromised production systems. The update worth noting now: NIST’s candidate list records that the HAWK submission team withdrew the proposal. Anthropic’s research report · NIST’s Round 3 candidate list.

What Claude found, and what it didn’t

A digital signature helps you verify who signed a file or message and whether it changed. It doesn’t hide the contents. HAWK was designed to do that job even in a world with capable quantum computers. Anthropic’s finding was about the cost of recovering a private signing key from that particular design. The team reported that recovery was substantially cheaper than previously estimated for the proposed parameters. Its analysis still found larger parameter sets impractical to attack. Anthropic’s research report.

The same report also described a separate result on a deliberately shortened version of AES-128: seven rounds instead of ten. That’s a tool researchers use to study how much safety margin a cipher has. It’s not a break of standard AES-128. Likewise, the HAWK result does not show that ML-DSA, SLH-DSA, or ML-KEM is broken. HAWK was a candidate, not one of NIST’s published standards. NIST’s post-quantum standards.

Why the withdrawal matters

In May 2026, NIST advanced HAWK and eight other candidates in a search for additional signature options. The withdrawal is a good example of why that process is public: it’s much easier to retire a proposal than to replace a signature scheme embedded in years of firmware, hardware, and signed software. NIST’s May announcement · NIST’s current candidate list.

There’s also a lesson about how the research happened. Anthropic describes a human-guided team of AI workers that reviewed prior work, pursued different ideas, and checked the result. The useful part isn’t that an AI declared a scheme weak. It’s that researchers could check the claim and disclose it to the people responsible for the design. The same standard should apply to any AI-generated finding a security team sees.

What platform teams can do now

Three illuminated checkpoints connect an older server rack to a newer one.
AI-generated illustration. Treat migration as a series of measured checkpoints, not a single cutover.

Start by separating two projects that often get lumped together: protecting data in transit and verifying what you trust. A pilot that changes how two endpoints agree on a secret doesn’t automatically upgrade the signatures on their certificates or on the software they run.

  1. Build an inventory with owners. For TLS, VPNs, SSH, code signing, and firmware updates, record the algorithm, where it is configured, and who can change it. Record key metadata, never private key material.
  2. Pick a pilot you can measure. Test a supported key-establishment option between endpoints you control and measure handshake failures, negotiation fallbacks, and latency. Separately, test whether your signing and verification tools can handle a standardized post-quantum signature.
  3. Practice the rollback before expanding. Old appliances, client libraries, and certificate tooling can break an otherwise sound rollout. Agree on a rollback trigger and track where legacy options remain enabled, with an owner and removal date.

NIST’s transition guidance and migration FAQ are good starting points. The HAWK finding doesn’t call for an emergency AES replacement. It does make the case for knowing where your cryptography lives and being able to change it without an outage.

Source notes

Follow these links to check the reporting and documentation behind this article.

Explore related topics