AI’s growing skill at mathematical and code-breaking problems has cryptographers asking whether it could eventually weaken the encryption behind online banking and secure messaging. But no currently deployed encryption algorithm has been broken by AI, and the experts quoted by Scientific American disagree on how much cause for alarm there is.
Watch Desk analysis
What happened
Scientific American reports that OpenAI recently put 722 AI-generated mathematical manuscripts online, withdrawing three within a day. None of the papers dealt with cryptography, the article notes. Even so, cryptographers are debating whether AI could help uncover mathematical shortcuts that weaken public-key encryption, which lets systems establish secure connections and verify digital identities.
The article points to a study that tested leading AI models against 191 code-breaking challenges. Its researchers say models’ cryptanalysis abilities are growing quickly, and that they found weaknesses that had not previously been spotted. Separately, Anthropic has reported that Claude found an improved attack on HAWK, a proposed digital-signature scheme designed to withstand quantum computers. These examples concern challenges and a proposed scheme, not a break of encryption protecting everyday internet traffic.
That distinction matters. The experts quoted by Scientific American differ on the broader risk: some warn that AI could help find shortcuts in widely used mathematical problems; others say there is no evidence of a breakthrough against the foundations of current public-key systems. The article says no currently deployed algorithm has been broken by AI.
Why it matters
Public-key cryptography helps protect connections to websites and the digital signatures used to verify software updates and cryptocurrency transactions. A major mathematical breakthrough could therefore have far-reaching consequences. But the path from solving a research challenge to breaking a deployed system is a long one, and the article’s evidence does not show that AI has crossed it.
The nearer-term question may be whether AI companies are testing their models against cryptographic problems, and what they share if they find something important. Researchers quoted in the article call for more transparency about what models have tried and what they have discovered. Preparing for quantum computers may help with some future risks, but the article notes that those risks are not identical to AI-assisted attacks.
Our read
This is a real security question, not a reason to change your passwords in a panic. The useful takeaway is that researchers are watching a capability grow, while the headline outcome, AI breaking encryption in widespread use, has not happened. We would like AI labs to explain what they are testing and to work with cryptographers on responsible disclosure. “Trust us, nothing happened” is not a particularly sturdy security plan.
What to watch
- Whether AI researchers publish methods and results from cryptanalysis tests.
- Whether further work identifies weaknesses in widely deployed systems, rather than experimental challenges.
- How cryptography standards and post-quantum preparations respond to new findings.
Discussion spark: Should AI labs be expected to disclose when they test models against cryptographic systems, even if they find no break, or could that disclosure create more risk than reassurance?
Sources and evidence
- Can AI break encryption? OpenAI’s new math has cryptographers divided – Scientific American (15 October 2026, 14:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.