Anthropic Claude Mythos Cryptanalysis Results

Anthropic's unreleased advanced model, Claude Mythos, has successfully produced two distinct cryptanalysis results: a new key recovery algorithm for the HAWK signature scheme and a modest improvement to an existing attack on reduced-round AES. These results demonstrate that advanced AI models can now understand, synthesize, and extend existing cryptanalytic research without detailed human intervention.

Impact on the HAWK Signature Scheme

The attack on HAWK is a significant result because it reduces the security margin of a proposed post-quantum-safe signature scheme. HAWK is based on the module Lattice Isomorphism Problem (module-LIP) and was being evaluated for future standardization.

Key takeaways regarding the HAWK attack include:

  • Security Reduction: The attack does not break HAWK in a practical "sci-fi" sense but roughly halves the number of security bits, meaning the attack remains exponential in time. While theoretically fixable by doubling key sizes, this would undermine HAWK's primary advantage: its efficiency.
  • Practical Demonstration: The attack produced functional code that successfully recovered keys against a weakened "challenge instance" of HAWK in a few hours of wall-clock time.
  • Methodology: The attack did not invent new mathematics. Instead, it synthesized and extended well-known, existing tools more thoroughly than previous human researchers had.

Analysis of the Reduced-Round AES Attack

While the result involving AES is technically a cryptanalytic progress, it does not pose a practical threat to the AES standard. The attack targets a 7-round variant of AES, whereas the full deployed version uses 10, 12, or 14 rounds.

The AES result is characterized by the following:

  • Marginal Improvement: The attack is a modest constant-factor improvement over work first published in 2013.
  • Impracticality: The attack requires $2^{89}$ cipher operations and $2^{105}$ chosen-plaintext encryptions, making it entirely impractical for real-world application.
  • Theoretical Nature: Because the operation count is so high, the result remains an on-paper analysis rather than a runnable attack.

AI Research Methodology and Verifiability

Anthropic's research process indicates that these results were achieved not through a team of domain experts tuning the AI, but by providing the model with general goals and allowing it to iterate. Prompts used were often simple, such as instructing the model to "do a breakthrough."

The Verifiability Bottleneck

As AI models generate more mathematical and cryptanalytic results, the bottleneck has shifted from discovery to verification.

  • Easy Verification: Full attacks, such as the one on HAWK, are easily verified by running the provided code against a challenge instance.
  • Difficult Verification: Subtle improvements, such as the AES result, require complex human review or formally-verifiable proofs (e.g., using the Lean language). However, these proofs are highly sensitive to how the theorem is formulated, still necessitating human expert oversight.

Implications for Cryptography and AI Capability

For Cryptography Users

Symmetric ciphers (like AES) remain robust due to their "messy" design, which is intentionally difficult to untangle. Public-key cryptography is more vulnerable because it relies on a small set of conjectured "hard problems" (e.g., RSA, lattice problems). The current transition to post-quantum algorithms provides an ideal window for AI-driven cryptanalysis to help identify and harden these new standards before they are widely deployed.

For AI Development

These results challenge two common narratives about current LLMs:

  1. Not "Glorified Autocomplete": The ability to synthesize existing tools into a new, functioning attack demonstrates a level of intelligence and capability beyond simple pattern matching.
  2. Not Yet AGI: Despite these breakthroughs, models still have a sharp "drop-off" point where they transition from being helpful to clueless, indicating that super-intelligence has not yet been reached.

Community Perspectives

Discussion among researchers and observers highlights a growing concern regarding "crypto-post-slop"—the potential for AI to generate a high volume of marginal, clickbait-style cryptanalytic results that claim to "break" standards while relying on impractical assumptions.

Regarding the availability of the model, some users note that while Claude Mythos remains unreleased to the general public, its capabilities are partially available through filtered versions like "Fable," which may downgrade the model's performance when cybersecurity-related queries are detected.

Sources