Discovering Cryptographic Weaknesses with Claude

Original Article ↗ Hacker News Discussion ↗

Anthropic’s claim that its Claude-based “Mythos” system found new cryptanalytic improvements on reduced-round AES and the HAWK post-quantum signature scheme has prompted debate over both the technical significance and the framing of the results. Commenters generally note that the attacks do not endanger real-world cryptography but see them as an early example of LLMs augmenting specialist research by searching large solution spaces, while questioning reproducibility, publication spin, and the $100k API cost per result. The thread also raises broader concerns about overreliance on immature post-quantum schemes, emerging inequalities in access to powerful models, and whether “prompt engineering” is a real skill versus just clear problem specification.

Scope and significance of the crypto results

  • AES work is on 7‑round AES‑128 with chosen‑plaintext; full 10‑round AES in real systems is not impacted.
  • The improvement over prior 7‑round attacks is described as marginal (a few bits of security) and part of a long chain of incremental attacks.
  • The HAWK result halves effective key strength for a simplified HAWK‑256 variant via an automorphism in its lattice; this may extend to HAWK‑512 but is not yet real‑world‑critical since HAWK isn’t deployed.
  • Several commenters emphasize this is standard cryptanalytic practice on reduced‑round / chosen‑plaintext models, not a practical break of deployed systems.

Post‑quantum crypto and security concerns

  • Some are uneasy about relying solely on PQC schemes and prefer hybrid (classical+PQC) modes; others counter that hybrids add complexity and overhead.
  • Debate over quantum timelines: some think serious quantum threats are still far off; others note major companies are already planning to deprecate pre‑quantum crypto.
  • There is skepticism about “LLM‑found PQC weaknesses” being used to push or resist PQC adoption.

Nature of the AI contribution

  • Many view the attacks as AI‑assisted refinements of existing ideas, not wholly new concepts.
  • Several liken this to LLMs finding counterexamples in math: heavy search with decent heuristics over known theory.
  • A key open question is reproducibility: how often does the system find the “good path,” and how much is stochastic luck vs multi‑agent collaboration.

Cost, access, and inequality

  • Each result reportedly cost around $100k in API usage over ~a week; some see this as evidence of an emerging “tech aristocracy.”
  • Others argue costs will fall as with early personal computers, and that internal teams naturally enjoy higher throughput than public users.
  • Comparisons are drawn to cheaper non‑US models, with disagreement over whether their lack of similar public results implies lower capability or just different priorities/secrecy.

Prompting, tools, and “AI usage” as a skill

  • The prompts used are informal, repetitive, and typo‑filled, leading some to downplay “prompt engineering” as a discipline.
  • Others argue the real skill lies in scoping tasks, managing context, noticing when the model drifts or hallucinates, and treating it like a junior collaborator.
  • There is debate over elaborate agent setups (skills, config files) versus simple tool access with light steering; experiences differ on whether extra structure helps or hurts.

Framing, PR, and research process

  • Several commenters feel the blog’s opening overstates novelty, given that the possibility of such weaknesses was already known.
  • Concerns about publication bias and marketing: many failed runs are invisible, while successful runs are heavily publicized.
  • Others defend this as legitimate science: checking the model’s claim, writing a paper, and subjecting it to community scrutiny is the normal process.
  • Some worry about over‑interpreting these results as “AI is better than cryptographers,” rather than “AI can accelerate incremental work within human‑defined frameworks.”

Broader implications for math and open problems

  • There is a concern that AI will make some problems trivial and others feel “hardened,” discouraging human effort because “even AI couldn’t solve it.”
  • Others think humans will continue working on what AI cannot do, potentially using AI as a tool to explore simple missed ideas or generate candidate constructions.