Going through a bunch of old audits to build a Solana program audit skill for agents. It's already finding > 60% of auditor findings with some very minor false positives (typically from not understanding intentional design decisions) tested with opus 4.5