OpenAI GPT–6 Astra breaks Enigma message that has resisted solution since 2005

Original Article ↗ Hacker News Discussion ↗

An experimental OpenAI model, GPT‑6 “Astra,” was used to crack a World War II–era Enigma cipher that had resisted decryption since 2005, prompting debate over how impressive the feat really is. Commenters argue over whether this is genuine progress in AI-assisted cryptanalysis or mostly a publicity stunt built on existing tools, human steering, and vast compute, with some pointing out that other modern models can likely do the same. The exchange broadens into questions about AI’s ability to tackle genuinely novel problems, its impact on human expertise and status, and whether current advances represent a transformative shift or an overhyped bubble.

Perceived significance of the break

  • Some see this as mostly PR: obscure recreational problem, likely solvable with enough compute and existing techniques, not a deep scientific breakthrough.
  • Others argue it is genuinely impressive: this specific message resisted experts for ~20 years; having an LLM autonomously pick it, build tools, and crack it is a strong capability signal.
  • Several people complain about “goalpost moving”: every time models do something new, critics declare that thing unimportant or “not real intelligence.”

How the message was broken (per discussion)

  • The Enigma message was unusual: different key from other traffic, transcription errors, and a rare left-rotor turnover around character 72, which breaks standard crib-based shortcuts and forces more brute-force search.
  • Astra reportedly:
    • Chose this specific unsolved message as “promising.”
    • Noted a related message and used the repeated place name “ROSENOW ROSENOW” as a crib.
    • Wrote Enigma and bombe simulators (Python/C++), then searched keyspace until a German-looking plaintext validated via re‑encryption.

Human vs AI contribution

  • Article claims Astra did the cryptanalysis “entirely on its own” after being asked only to try unsolved messages on a specific site.
  • Some readers still describe this as human–AI “collaboration” or “tool use”; others say that’s overstating the human role here.
  • Debate over credit: is the achievement mainly the model’s, the prompter’s, or the prior human cryptanalysis encoded in training data?

Replications and other models

  • One commenter mistakenly “decoded” the already‑decrypted plaintext German; others pointed out the error.
  • A later experiment with Gemini 3.8 (with tools/agents) reportedly reproduced a full break in ~45 minutes by discovering open‑source Enigma code, compiling it, and running large key searches.
  • This is used to argue Astra isn’t uniquely capable; similar frontier models with tool use likely can do the same.

AI capabilities, creativity, and limits

  • Long subthread on whether this shows genuine problem‑solving versus mere pattern‑matching over human work:
    • Pro‑AI side: pattern‑matching is part of intelligence; building tools and choosing strategies from text is nontrivial.
    • Skeptical side: models are recombining known methods from training data; they lack “imagination” or world models, and struggle on genuinely novel problems.
  • Related philosophical debate: intelligence vs consciousness, whether human intelligence is still “special,” and emotional reactions to feeling intellectually displaced.

Cryptography implications

  • Several note that breaking weak historical ciphers is very different from breaking modern symmetric or public‑key crypto; no evidence current LLMs threaten strong cryptography.
  • Some speculate LLMs might eventually design new crypto, but emphasize that designing secure schemes is much harder than attacking broken ones and full automation would require reliable iterative validation loops.

Hype, business models, and social impact

  • Many see this as part of a pattern of AI labs chasing media‑friendly stunts (unsolved ciphers, math problems) to build hype, drive valuations, and frame benchmarks.
  • Others defend such demos as valuable evaluations of model progress when standard benchmarks are saturated.
  • Broader worries include: commodification of expertise, job displacement, concentration of power/wealth, environmental cost of compute, and the possibility that intelligence has become “a thing you can buy with electricity.”