A thousand fewer paths to root
OpenAI reports that its agent pipeline found and exploited 948 distinct Linux local privilege-escalation vulnerabilities in a mid-October scan, at $1.21M in API credits. The company says it generated candidate patches and is working with maintainers to report them; these figures are its own, and candidate patches are not necessarily merged or independently verified.
The pipeline separates discovery, reproducing faults, exploit development, patch generation, and human review. A notable failure mode: in an early run, 20 of 200 apparent successes used a different vulnerability than the one assigned, prompting an independent check of each exploit path. HN readers debate whether automated auditing changes the case for memory-safe code; others ask for comparable tests on smaller kernels.