Anthropic's Mythos found a better attack on HAWK in 60 hours — here's what the $100k cryptography run means for post-quantum security

Anthropic ran Claude Mythos Preview against two cryptographic targets for 60 hours last month. The model found a better attack on HAWK, a post-quantum signature scheme that human cryptographers reviewed for more than two years. Total API cost: roughly $100,000. Neither result affects systems in production today, but the capability and cost structure are worth understanding.

What Mythos actually found

HAWK (Hawk's Another Way to sign with Keccak) is a lattice-based signature scheme submitted to NIST's post-quantum cryptography standardization process. Human cryptanalysts found an attack that required 2^59 operations. Mythos found one that needs 2^48 operations — 2,048 times faster. That's still impractical for real attacks, but it's a meaningful improvement over the human baseline.

The second target was a deliberately weakened 3-round AES variant. Mythos rediscovered a known differential attack. Anthropic included this as a positive control — the model should find it, and it did.

Neither finding compromises deployed cryptography. HAWK isn't standardized yet, and 3-round AES is a toy problem. The research demonstrates capability, not immediate risk.

The prompts that worked

Anthropic published the exact prompts they used, typos included. The repo is on GitHub. One excerpt:

"the models tend to think it is impossible to break these schemes, so we had to prompt them to ignore their safety training and focus on the mathematical properties."

That's the interesting part. Mythos needed explicit instructions to override its refusal patterns. The prompts tell the model to treat the problem as pure math, not as a request to "hack" something. The framing matters more than the technical depth.

The approach combined chain-of-thought reasoning with iterative refinement. Mythos generated candidate attacks, Anthropic's team ran verification scripts, then fed the results back into the context. Total wall time: 60 hours. Most of that was automated looping, not human review.

What $100k buys

Mythos Preview costs roughly $15 per million input tokens and $75 per million output tokens under Anthropic's enterprise pricing. A 60-hour run at $100k works out to about 1.3 billion tokens processed — reasonable for iterative cryptanalysis with long context windows.

That's expensive for a single research question, but it's cheap compared to a human cryptography team running the same analysis over months. HAWK's original review involved multiple researchers across years. Mythos compressed part of that search space into a weekend.

The cost structure inverts as models improve. GPT-5.6 and Fable 5 are both cheaper per token than Mythos Preview. If the capability generalizes, the price drops fast.

The post-quantum timing

NIST is finalizing its first batch of post-quantum standards right now. HAWK didn't make the initial cut, but it's still under review for future rounds. Anthropic's result won't disqualify HAWK — the attack is still impractical — but it changes the review calculus.

Post-quantum cryptography assumes that large-scale quantum computers will break RSA and ECC within the next 10-20 years. The replacement schemes rely on hard problems from lattice math, code-based crypto, and hash-based signatures. Human cryptanalysts have spent decades stress-testing these constructions.

Mythos is now part of that stress test. If a $100k model run can find better attacks than a multi-year human review, that's a forcing function. Future submissions will need to account for LLM-assisted cryptanalysis as part of the security model.

What this doesn't mean

Mythos did not "break encryption." The HAWK attack is 2,048 times faster than the previous best, but 2^48 operations is still infeasible. The 3-round AES result is a known differential attack on a toy variant. Real AES-128 uses 10 rounds and remains secure.

The research also doesn't prove that LLMs will automate all cryptanalysis. HAWK and 3-round AES are well-studied targets with published literature. Mythos synthesized existing techniques; it didn't invent a novel attack class. The next question is whether this generalizes to less-studied schemes.

Anthropic framed the work carefully. The title is "Discovering cryptographic weaknesses," not "Breaking cryptography." The distinction matters.

Where this lands for agent deployments

We don't deploy cryptanalysis agents. Most VioX clients run voice agents on Twilio, CRM tooling on Supabase, or multi-tenant platforms on Railway. The HAWK result doesn't change that stack.

But the cost and capability structure does. If Mythos can compress a multi-year cryptography review into 60 hours at $100k, that same pattern applies to other research-heavy domains — legal discovery, actuarial modeling, clinical trial design. The bottleneck shifts from "can the model do this" to "can we frame the problem so the model attempts it."

The prompt engineering is the hard part. Anthropic's team had to override refusal training and structure the task as pure math. That's closer to our work on evals and goal loops than to the conversational agents most teams deploy.

Cryptography is a clean testbed because verification is deterministic. You run the attack, count the operations, check the output. Most agentic domains don't have that property. The HAWK result proves capability; translating that to messier domains is the open problem.