A Dummy's Guide to Agentic Exploit Generation
summary
Agentic pipelines wire LLMs into an exploit workflow by combining deterministic tools (scrapers, static analysis, kprobe) with an orchestration agent that maintains state, uses templates and modular code generation, and enforces iterative feedback. Pure LLM runs on 58 bugs produce 4 working PoCs (many hallucinate or “cheat”), while the agentic system yields 21 roots, 16 previously unknown CVEs. Key specifics: preserve memory of failures, pull symbols from the target, add runtime probes, and isolate small code functions for repeatable compilation and verification.
AI-written summary · about
chapters (14)
- 0:00Agentic exploit generation intro
- 1:13Manual kernel exploit steps
- 5:06Use-after-free example
- 9:02LLMs fail at exploit gen
- 10:06AI-driven exploit research state
- 12:56Exploit gen key problems
- 15:23Agentic vs traditional pipelines
- 17:12Consumption and root cause analysis
- 20:02Planning and code generation
- 23:11Compilation and verification challenges
- 24:34Verifying exploits without false positives
- 26:18Feedback loops and iteration
- 27:43Results and real-world testing
- 29:50Lessons learned and best practices
related talks
15mClearAgent: Agentic Binary Analysis for Effective Vulnerability Detection
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
29mBuilding a Hackbot: How Ethiack's Hackian Finds Real Vulnerabilities
20mAutomated Exploit Generation for Node.js Packages
47mAgentic Testing: Your Skilled Human Tester
13mThe Automation of Exploits: Will Eatherton on AI-Driven Security and the End of Obscurity
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
39mThe AI Arms Race in Vulnerability Management, Who’s Winning?