What's this about?
HN commenters read AISLE's homepage and repository and worked out the method. It is AI-guided fuzzing built on libFuzzer or AFL: have the LLM identify interesting functions to test, have it generate the fuzzing harnesses, then have it triage the fuzzer's findings at a high level. And one commenter found the scanner's default model in the code: gpt-5.4-nano — not a frontier model, and not even a large one.
Why does that change the story?
Because the headline reads as "a small company's AI beat OpenAI's and Anthropic's", and the mechanism is different: a cheap model wrapped in a good harness beat two frontier models wrapped in whatever Codex Security and Mythos use. Fuzzing is a proven technique that generates *executable evidence* — a crashing input, not a claim — which is why the findings survived expert review. The model's job was to point the fuzzer somewhere useful and sort the results, and a nano model can do that.
Why does this matter to you?
- It is the strongest empirical case this month for the argument three separate papers have made: the harness dominates the model.
- It arrives the same day as FrontierHarness measuring a 17× cost spread across harnesses running the identical model.
- The practical read for security tooling: pairing an LLM with a deterministic tool that produces verifiable artifacts beats asking a bigger model to reason about code.