In this keynote address, Linux kernel maintainer Greg Kroah-Hartman analyzes the impact of Large Language Models (LLMs) and automated AI agents on Linux kernel development and security workflows. Address topics include the escalation of kernel Vulnerability Disclosures (CVEs) from 50 per week to 33 per day, a detailed post-mortem of Anthropic’s "Mythos" vulnerability report (where 79 claimed bugs yielded only 10 actionable bug fixes), and empirical observations showing a 50% failure/false-positive rate in LLM-generated kernel patches. The speaker frames LLMs not as intelligent reasoning engines, but as deep static pattern matchers analogous to Coccinelle or automated fuzzers. Maintainer strategies are presented to mitigate AI-generated patch spam, enforce local model execution to prevent data leaks, document subsystem threat models, push back against AI "meat proxies," and handle incoming security reports without operational panic.
Key Highlights & Timestamps
00:00 CVE Processing Capacity: Linux kernel CVE volume expanded from 50 per week to 33 per day, supported by scaled CNA infrastructure and automated tooling.
02:48 Analysis of Anthropic Mythos Report: A headline report alleging 79 kernel vulnerabilities yielded only 20 actual fix candidates and 10 real bug fixes—representing approximately one hour of standard kernel development output.
11:05 Compressed Exploit Timelines: The traditional discover-to-exploit window has shifted from 63 days post-disclosure to -7 days pre-disclosure, driven by AI agents systematically chaining minor flaws.
13:08 Static Pattern Matching vs. Fuzzing: LLMs act as fuzzy static analysis tools like Coccinelle rather than intelligent entities; their output resembles legacy Coverity scanning waves or fuzzers that must be systematically remediated.
18:26 50% False Positive Rate in AI Patches: Empirical review of AI-generated kernel patches reveals a ~50% failure rate, including non-functional logic, invalid assumptions, and hallucinated security impacts.
22:15 Hallucinated Code & Pattern Errors: LLM patch submissions regularly exhibit nonsensical transformations (such as replacing mutex_unlock with mutex_destroy), over-commenting, and verbose change logs derived from outdated kernel commit histories.
28:30 Threat Modeling & Subsystem Defense: Maintainers are establishing explicit threat model documentation (e.g., trusting hardware, disregarding malicious mount images) to reject invalid AI bug reports upfront.
31:08 Safe Execution via Local Models: Open-source AI agents (such as Shashiko) should be executed strictly on local hardware; non-public kernel code or security bugs must never be uploaded to commercial cloud APIs due to data leakage risks.
34:20 Community Protocols & Review Strategies: Maintainers are urged to demand re-producers from submitters, reject raw AI text walls, strip unverified sign-off tags, and prioritize code deletion over unmaintained fixes.