Anthropic Claude Mythos Preview Discovers Cryptographic Flaws and Launches CryptanalysisBench
Anthropic demonstrated that Claude Mythos Preview can autonomously discover mathematical flaws in core cryptographic algorithms. The model halved the key strength of the post-quantum candidate HAWK in 60 hours and accelerated round-reduced Advanced Encryption Standard attacks by up to 800 times. Anthropic also released CryptanalysisBench in collaboration with academic institutions.
Impact: High
Why it matters
You can evaluate AI agent reasoning on complex mathematical and cryptographic tasks using CryptanalysisBench or adapt multi-agent Python and Sage execution harnesses for automated code auditing.
TL;DR
- 01Agentic harnesses combining LLMs with execution environments like SageMath can perform complex symbolic mathematical discovery.
- 02Post-quantum cryptographic candidates like HAWK require larger safety margins as AI automates cryptanalytic search.
- 03CryptanalysisBench offers a standardized benchmark to measure frontier model capabilities on cryptographic tasks.
Key facts
- HAWK Attack Compute Time
- 60 hours
- AES Attack Acceleration
- 200x - 800x
- API Cost Per Result
- ~$100,000
- Benchmark Name
- CryptanalysisBench
Autonomous Cryptanalysis in Agentic Harnesses
Anthropic paired Claude Mythos Preview with a multi-agent harness inspired by Claude Code. The setup orchestrated worker agents inside a sandboxed environment with access to Python, SageMath, and cryptographic literature. Operating semi-autonomously, the model conducted literature reviews, performed mathematical reasoning, and constructed automated verification pipelines to prove its findings.
Impact on HAWK and Advanced Encryption Standard
- HAWK Post-Quantum Signatures: Over 60 hours, Mythos discovered a non-trivial lattice automorphism, cutting effective key strength in half and requiring doubled key sizes to maintain security.
- Round-Reduced AES: Mythos eliminated a guess step in existing attack vectors, increasing execution speed by 200x to 800x.
- Research Spend: Developing each attack vector cost approximately $100,000 in API inference costs over a one-week testing period.
Benchmark Release
To benchmark model capabilities in mathematical and cryptanalytic reasoning, Anthropic released CryptanalysisBench alongside researchers from ETH Zurich, Tel Aviv University, and University of Haifa.
Try it in 2 minutes
# Evaluate agent reasoning on CryptanalysisBench
git clone https://github.com/cryptanalysisbench/benchmark.git
cd benchmark && pip install -r requirements.txt
python run_eval.py --model claude-3-5-sonnet-20241022 --tools sage,pythonbash
✓ When to use
- Benchmarking frontier LLM capabilities on complex symbolic logic and formal verification.
- Designing multi-agent sandbox harnesses with specialized execution tooling like Python and Sage.
✕ When NOT to use
- Panicking about immediate RSA or full AES-256 production vulnerability.
- Expecting standard chat prompt interfaces to solve high-order math without tool execution harnesses.
What to do today
- Review post-quantum signature schemes and key size specifications in upcoming deployments.
- Explore CryptanalysisBench if building domain-specific math or security agent harnesses.
What the community says
“In this case it's not actually that big a result, it's an incremental improvement on a series of previous results... improving the Derbez, Foque, Jean attack from EUROCRYPT 2013.”
“Keep in mind for Anthropic the goal for them isn't necessarily to solve a problem but to demonstrate/measure the innate capabilities of their model.”
Sources