Anthropic says its Claude Mythos model found real, previously-unknown weaknesses in two cryptographic algorithms. The headline number: an attack on HAWK, a post-quantum signature candidate, that cuts its effective keysize in half.
This is not a “the sky is falling” story. Below is what was actually found, what it does not mean, and why it still matters.
The News
Anthropic, working with academic partners including ETH Zurich and Tel Aviv University, had its Claude Mythos Preview model attack two targets:
- HAWK — a post-quantum digital signature scheme, a third-round candidate in NIST’s search for cryptography that resists quantum computers. Claude found a new attack that cuts HAWK’s effective keysize in half. For HAWK-256 specifically, the expected attack cost dropped from 2^64 to 2^38 — a huge reduction in the work needed to break it.
- A weakened 7-round version of AES-128 (out of the full 10 rounds) — a research variant used to study attack techniques, not the cipher used in production. Claude improved the best-known attack by 200 to 800 times, using a technique the researchers call “Möbius Bridge.”
The HAWK result took about 60 hours. The AES result took 3 days of Claude’s work plus several hundred hours of human validation.
What It Does Not Mean
Both headlines undersell an important caveat:
- Full AES is untouched. The attack works on a deliberately weakened 7-round test version, not the 10-round cipher your data is actually encrypted with today.
- HAWK is not yet a deployed standard. It’s a candidate still going through NIST’s evaluation process.
Anthropic states this directly: “Neither of these results has a practical impact on today’s computer systems; no production software will have to change as a result.”
It Was Not Fully Autonomous
This is the part most headlines skip. Claude did not sit down and solve these problems on its own:
- For HAWK, a researcher gave occasional guidance and nontechnical direction to keep the model on track.
- For AES, a researcher built the initial scaffolding, then sent three substantive prompts over three days to redirect Claude’s focus when it stalled.
The researchers’ own framing: models “tend to think it is impossible to solve, so they don’t try” — they needed real prompting to keep going instead of giving up or settling for easy, unpublishable findings.
The Cost
Each result cost roughly $100,000 in API usage to produce. This was a research effort with dedicated human oversight, not something that happens by accident or at low cost.
What Anthropic Built Alongside This
The collaboration also produced CryptanalysisBench, a benchmark that packages many cryptographic ciphers together so researchers can measure how well language models perform at cryptanalysis. This matters more than either single result: it turns “can AI break crypto?” into something you can actually track over time, model generation to model generation.
What It Means for Developers
1. This is a capability signal, not an incident. No system you run today needs to change because of this. Treat it as evidence that frontier models are approaching expert-level cryptanalysis skill, not as a breach.
2. Watch CryptanalysisBench scores over time. As models improve on this benchmark, it becomes a leading indicator for which primitives deserve renewed scrutiny — before an attack targets something actually deployed.
3. Post-quantum candidates are exactly where this kind of pressure-testing should happen. HAWK is still in NIST’s evaluation pipeline. Finding weaknesses now, before standardization, is the system working as intended.
Quick Summary
| Question | Answer |
|---|---|
| Model used | Claude Mythos Preview |
| HAWK result | Effective keysize cut in half (attack cost 2^64 → 2^38) |
| AES result | 200-800x faster attack, on a weakened 7-round variant only |
| Fully autonomous? | No — required human guidance and redirection |
| Cost | ~$100,000 in API usage per result |
| Affects production systems today? | No |
FAQ
Does this mean AES is broken? No. The attack targets a deliberately weakened 7-round version of AES-128 used for research, not the full 10-round cipher that protects real systems.
Is HAWK unsafe to use? HAWK is still a candidate in NIST’s post-quantum standardization process, not a deployed standard. This kind of attack is exactly what that evaluation process exists to catch.
Did Claude do this on its own? No. Both results required a human researcher to steer the model — encouraging it to keep trying and redirecting its focus — over multiple days.