Frontier AI labs have begun quietly testing whether their newest math-capable models can break important cryptographic protocols and primitives. This is not speculation. Scott Aaronson writes that frontier AI labs have begun quietly checking whether their newest math-capable models can break important cryptographic primitives, after OpenAI's October 6 release included claimed proofs of the Unique Games Conjecture and L=BPL.
The evidence for secrecy is the silence. OpenAI disclosed 372 new mathematical results its unreleased model generated, a spokesperson says nearly all came from one prompt to a single AI agent, with 'some' needing multiple attempts. The approach costs far less than the 10,000-agent swarm that solved Navier–Stokes; OpenAI withheld prompts and per-problem compute times despite advisory-group calls for full disclosure. More telling: Aaronson noted that among the 722 mathematical results recently published by OpenAI, there is a striking lack of breakthroughs in cryptography, and the US government has censored academic results on quantum cryptanalysis.
That absence is a tell. When you're generating 372 mathematical proofs across open problems, the absence of cryptography results isn't random. It suggests labs have found things they're not ready to disclose, either because the results are startling, or because national security handlers have asked them to hold.
The math itself is stunning. OpenAI's internal model solved ~5% of about 8,000 attempted open problems at roughly 3 hours of compute each, with Lean certificates attached but natural-language writeups that mathematicians describe as 'written by someone on psychedelics.' A 5% success rate on hard open problems, run at machine speed, across 8,000 attempts: that's a different regime entirely. If this model can crack difficult math in hours, then applying it to cryptanalysis, a problem space where computational difficulty is the entire point, is obvious.
There's precedent for this kind of work staying hidden. In July, Anthropic's Mythos AI systems were able to discover previously unknown weaknesses in multiple cryptographic algorithms previously thought resilient, including the HAWK post-quantum digital signature scheme and a reduced-round version of AES. The lab published this work, but what didn't they publish? What stayed in the lab? While Anthropic emphasized that neither discovery affects production banking systems today, nor does the work compromise full-strength AES, the research nonetheless demonstrates that frontier AI models are becoming capable of performing sophisticated cryptanalysis previously requiring years of specialized human research.
The asymmetry is the story. Labs can now run cryptanalysis at machine speed and at scale. The public has seen samples of what they're willing to release. The secret work is the thing we're not seeing. And the government, presumably, is very interested in what stays hidden.
Aaronson's read is direct: labs are running this work, and they're not saying so. The absence of crypto results from OpenAI's latest dump, the known precedent from Anthropic, the mention that rumors suggest that OpenAI's most recent release of mathematical findings is the first of three planned batches, all of this points to work being staged and filtered. What's coming in the other batches is the real question.
Top comments (0)