AI for Web3 Security: Real-World Use Cases — Dr. Stefan Beyer
summary
Company uses AI to speed audits with a co-pilot model that runs LLMs (GPT‑4 class) for surface scanning, code comprehension, fuzzing and report drafting, freeing humans for business‑logic, economic and game‑theory analysis. Internal metrics: 52 researchers, 600+ audits since 2017. Research shows ~96% precision when AI flags vulnerabilities but ~40% of issues are missed; fuzzing yields ~90% precision. Case: a Cosmos SDK audit found one tool-detectable unbounded execution and one complex business‑logic bug AI did not find, so LLMs run locally under NDA.
AI-written summary · about
chapters (9)
- 0:00AI in web3 security intro
- 1:02AI accessibility and coding impact
- 3:10Autonomous security with AI
- 4:56AI for vulnerability detection
- 5:59Human vs AI in audits
- 7:05Co-pilot model for audits
- 9:10Optimizing LLM for audits
- 11:39Local LLM and data protection
- 12:41Operational security risks beyond code
related talks
31mAI for Web3 Security
57mAI Applications Security Puzzle
55mAI-First Testing
28mAI for Everything After Code
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
20mSecurity, and AI... so happy together?
47mAI and the Future of Software Development
42mAI for security - friend or foe?