On this page
AI on defense: audits, tracing, and the arms race
The same tools cut both ways — which is the actual shape of AI in crypto security: an arms race, not a one-sided breakthrough. On defense, are already useful review assistants — summarizing code, flagging suspicious approvals, drafting smart-contract audit findings — and the US government is betting that AI can find and patch real vulnerabilities: DARPA's AI Cyber Challenge ran AI systems against critical open-source code from 2024, with finals at DEF CON 2025. On the tracing side, firms were machine learning shops before it was fashionable — clustering heuristics and anomaly detection are what the Tracing unit described, and better models make surveillance cheaper every year (the privacy-versus-tracing tension sharpens accordingly).
The honest caveats: LLMs hallucinate — they will confidently write a contract that contains a reentrancy bug, or “audit” one and miss it, and attackers run the same tools you do. AI shifts the cost curve for finding bugs on both sides; it does not change which side wins. What still decides outcomes is process — audits, bug bounties, battle-tested patterns — which is why the advice in the Smart Contracts unit (don't trust vibes, trust verification) survives contact with AI unchanged.