AI meets Cryptography 3: What AI Found in Bron Labs's bron-crypto
12 hours ago
- Bron-crypto's Lindell17 DKG encrypted share computation swapped operands, corrupting threshold ECDSA key shares and causing signing failures.
- Poseidon hash.Hash adapter broke Go's contract by resetting state on each Write and misusing Sum for input, resulting in wrong hashes for streaming.
- IsOnCurve checks for k256, Pallas, and Vesta were inverted, rejecting valid points and accepting invalid ones, enabling invalid-curve attacks.
- IsZero in BLS12-381 G2 field element incorrectly checked IsOne, causing identity detection errors but with low practical impact.
- The set of bugs found by LLMs is stable across runs, with duplicates increasing, suggesting repeated runs yield diminishing returns.
- Optimizing the harness around the model is key to discovering new vulnerabilities, not just re-running the same setup.
- Maintaining an effective LLM audit setup requires continuous iteration due to evolving model strengths and responses.