Theoretical computer scientist Scott Aaronson has claimed, citing sources, that AI laboratories have quietly begun testing whether their models can break important cryptographic protocols. He published the claim on his blog, Shtetl-Optimized, as flagged by Techmeme.
Cryptography protects everything from bank transfers to private messages. If an AI model could crack widely used protocols, the security of the entire internet would be at risk. Aaronson's warning matters because he is a credible voice in theoretical computer science, not an anonymous forum poster.
Aaronson's cryptography warning
The cryptography claim arrived in a long post dominated by another story: OpenAI's release of 372 mathematical results, including a proof of Subhash Khot's Unique Games Conjecture. Aaronson wrote that the proofs appear to have a Lean certificate, but that no human has yet understood almost any of them.
His wife, complexity theorist Dana Moshkovitz, has worked toward proving the Unique Games Conjecture for her whole career. In messages Aaronson shared, she described the AI proof as bizarre, badly written and impossible to read without AI help, inventing what she called a completely new code with a noise test.
AI math breakthroughs and two publishing models
Aaronson said the model behind the mathematical results was not a special system burning millions of dollars of compute, but simply the latest internal OpenAI model. It might reach paying ChatGPT customers within a couple of months, depending on OpenAI's safety board.
He contrasted two ways of publishing AI breakthroughs. OpenAI posted its undigested solutions directly, setting off a race among humans to understand them, while Anthropic gave two researchers the chance to write a digested version in exchange for compensation. Aaronson asked readers which model they preferred, without endorsing either.
