Anthropic researchers recently used Claude Mythos to find genuine mathematical weaknesses in two cryptographic systems: the HAWK signature scheme and a reduced-round variant of AES-128. Neither finding threatens any system in production today. What makes the research worth reading closely isn't the vulnerabilities themselves — it's what the researchers had to do to get the model there.
This article covers what Claude actually found, why 60 hours and roughly $100,000 in compute went into research with no practical security impact, and what the published prompts reveal about the current limits of AI-driven mathematical research.
What was actually found — and why it doesn't matter yet
Cryptanalysis is the practice of studying cryptographic systems to find weaknesses in the mathematical assumptions or implementations they rely on. HAWK is a newer signature scheme being evaluated as part of the broader shift toward post-quantum cryptography — algorithms designed to remain secure even against future quantum computers. AES-128 is one of the most widely deployed encryption standards in the world, but the version Claude analyzed was deliberately weakened — a reduced-round variant, meaning fewer of the standard's internal transformation steps were applied, making it an easier academic target than the real thing.
Finding a weakness in a reduced-round toy version of AES is a well-established genre of cryptography research; it doesn't imply anything about full AES-128, which remains unbroken. The HAWK finding matters more because it's a newer, less battle-tested scheme, but even there, the researchers were explicit that the discovery carries no practical security implications for current systems. The value of both findings is academic: they demonstrate a capability, not a threat.
The real finding is in the prompts
Anthropic published the prompts used during the research, and the detail that stands out is procedural rather than mathematical: researchers needed persistent encouragement to prevent the model from abandoning difficult problems partway through. That's a strikingly human failure mode to see documented in a frontier model's research process — not "the model got the math wrong," but "the model needed to be talked into not giving up."
The bottleneck in AI-driven mathematical research isn't always capability — sometimes it's persistence, and persistence turns out to be something researchers still have to manage by hand.
This matters more than it might first appear. A 60-hour, $100,000 research run is only economically defensible if someone is confident the model will actually converge on a result rather than stall out on the ninth hour of a plausible-looking dead end. If sustained problem-solving still requires a human periodically re-motivating the model, that's a real operational cost that doesn't show up in a benchmark score.
CryptanalysisBench and what comes next
The research produced a longer-lived artifact than either individual finding: CryptanalysisBench, a new evaluation framework built with ETH Zurich, Tel Aviv University, and the University of Haifa specifically to assess LLM capabilities in cryptanalysis research. Purpose-built benchmarks like this matter because generic coding or math benchmarks don't capture the specific skill of sustained, open-ended mathematical exploration — the skill this research was actually testing.
Having an academically validated benchmark for this capability also gives the field a way to track whether "needing persistent encouragement" is a temporary limitation of current models or a more durable feature of how these systems approach open-ended research. That's a more useful long-term signal than any single vulnerability disclosure.
What this means for builders
Teams considering AI-assisted research or long-horizon technical exploration should treat this as a data point on cost and reliability, not just capability. A model that can find a genuine mathematical weakness given 60 hours of compute and active human steering is different from a model that can be handed a hard problem and left alone — and right now, the gap between those two is where most of the real engineering cost lives.
If you're building agentic workflows meant to run unattended for hours, the "needed encouragement to not give up" detail is worth taking seriously as a design constraint, not a footnote. Persistence-checking and stall-detection may need to be explicit parts of the harness, not an assumption baked into the model.
Conclusion
The headline numbers — 60 hours, $100,000, real cryptographic weaknesses — make for a striking research story, but the more durable lesson is procedural: current frontier models can do genuine novel mathematical research, but doing so still requires active human management of the model's tendency to give up. Until that changes, the cost of AI-driven research isn't just compute — it's the researcher's ongoing attention.