AI for Web3 Security
summary
Oak Security runs 52 researchers, 600+ audits and 7,000 issues and reports that LLMs find standardized bugs fast but with terrible signal‑to‑noise: a benchmark where OpenAI 03 detects a Linux kernel use‑after‑free only in 8 out of 100 preformatted runs and drops to one run without formatting. An AI pentester files about 1,600 reports with ~5% critical and ~45% pending. Formatting yields an 8x performance boost; all AI findings require human validation and layered defenses.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 3:23AI cybersecurity reality
- 8:45AI pentester analysis
- 12:23LLM text generation limits
- 15:49LLM basics and challenges
- 19:51Oak Security LLM audits
- 22:48AI in development lifecycle
- 23:15Web3 operational security
- 27:26Q&A smart contract security
- 28:14LLMs attacker vs defender
- 29:50LLMs impact on testing
related talks
14mAI for Web3 Security: Real-World Use Cases — Dr. Stefan Beyer
57mAI Applications Security Puzzle
20mSecurity, and AI... so happy together?
51mAI for the People
39mThe AI Arms Race in Vulnerability Management, Who’s Winning?
44mTL;DR: Applying AI to Security
48mAI Impact on Cybersecurity and Privacy
24mAI Is Writing Your Bug Reports. Can You Tell?