OpenAI's agent pipeline found 948 Linux root exploits—20 early "successes" used the wrong bug
shoumikgoswami · reddit · 2026-11-05
OpenAI detailed an autonomous security pipeline run against the Linux kernel in mid-October: over 10 days, agents scanned 2.8M lines of security-relevant code, turned 2,308 candidates into 1,724 reproduced vulnerabilities (reproduction in 30 hours), ending with 948 working local privilege-escalation exploits plus candidate patches, all human-reviewed. Key failure mode: in an early run, 20 of 200 reported exploits got root via a completely different bug than assigned, forcing OpenAI to add independent verification—verify the path, not just the result, when agents grade their own work. A 103-technique exploit library helped the weaker model but not the stronger (+1.5 pts), so the scaffolding was deleted. Caveats: numbers are OpenAI's own; maintainers had independently patched 61 of the bugs during the scan.
More from coding & agent
- Coolify TLS cert rotation works, but CA API buckles under eMudhra fallout rush — tristanbob · 2026-11-14
- Delta editor adds Claude Agent via ACP, works with existing Claude subscriptions — dean_rie · 2026-11-14
- Claude Code 2.1.324 is about to be released — ClaudeCodeLog · 2026-11-14
- Matthew Berman open-sources Scarlet, an AI video editing agent that speeds up his workflow — MatthewBerman · 2026-11-14
- kirillzzy Launches Alien, an AI Agent Pitched as the Most Private Yet — Scobleizer · 2026-11-14
- 3DSceneBench: coding agents rebuild livable 3D rooms but top functionality score is just 42.1 — yizhu59 · 2026-11-14