Bankless AI Finds 70% of Smart Contract Exploits | Alpin Yukseloglu
22 snips
Mar 5, 2026 
Alpin Yukseloglu, investment and research partner at Paradigm and co-author of EVMBench, studies AI-driven smart contract security. He discusses AI moving from ~12% to 70% exploit detection, how verifiability makes crypto an ideal training ground, and the unfolding 24/7 blackhat vs whitehat arms race. He also explains EVMBench’s detect/patch/exploit benchmark and why labs may soon prioritize crypto data.
AI Snips
Chapters
Transcript
Episode notes
AI Jumped From Novice To Highly Effective Smart Contract Auditors
- AI auditor capabilities rose from ~12–13% to over 70% exploit-finding on EVMBench within months.
- Alpin credits rapid model improvements (e.g., GPT-5.3 codecs) and fast capability absorption during development of the benchmark.
Superhuman Auditors Could Break Assumptions This Year
- Superhuman AI auditors may arrive within months, breaking current security assumptions for high-value contracts.
- Alpin expects models to outpace median engineers first, then best auditors, forcing industry hardening by year-end.
Move From Worry To Action By Building Defenses
- Stop theorizing about distant singularities and act: join experiments and build defensive tools now.
- Alpin recommends getting into the trenches with model labs and contributing practical integrations to shape outcomes.