Keynote: Linux in the Land of LLMs
summary
LLMs detect bugs by pattern matching — Anthropic’s Mythos style scans reproduce old-vulnerability patterns against current code — and that drives CVE churn from ~12–13 to 14–15 CVEs per day. Models lie and are wrong roughly 25-33% false positive rate, so the kernel requires bug reports to include an initial patch, which removes about a third of spurious reports. Recommendation: run models locally, use open-source threat-model generators (Alpha Omega), reuse kernel review prompts and local agents, and expect an intense 18-month bug-fix push.
AI-written summary · about
chapters (7)
- 0:00LLMs in Linux kernel
- 3:15LLMs for security bugs
- 6:27False positives and review
- 8:58Kernel LLM security adoption
- 11:46Practical steps to use LLMs locally
- 14:17Local LLM tools and agents
- 15:40Q&A
related talks
36mUsing AI to Discover Silently Patched Vulnerabilities in Open...
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
32mFuture Open Source LLM Kill Chains
11mCan LLMs Understand the Linux Kernel?
46mHow Large Language Models Are Reshaping the Cybersecurity Landscape
29mIdentifying and Assigning AI Model Vulnerabilities
57mVulnerability Root Cause Mapping with CWE
47mHarnessing LLMs for Next-Level Security Analysis