OpenAI says an autonomous AI cybersecurity pipeline produced working exploits for 948 distinct Linux kernel local privilege-escalation vulnerabilities and generated candidate patches in a mid-October research run that cost $1.21 million in API credits.
The company detailed the experiment in its November 4 research report, which used an upcoming cyber-focused model in a controlled environment.
The pipeline began with 2,308 candidate findings and dynamically reproduced 1,724. Exploit-development agents produced 1,169 working local privilege-escalation exploits, then variant analysis grouped them into 948 distinct vulnerabilities. The work also produced 291 proof-of-concept exploits for information-leak flaws. The privilege-escalation bugs could allow an unprivileged local user to gain root access.
OpenAI said the pipeline ran against a stock desktop operating system in isolated virtual machines with standard Linux protections enabled. Its five stages covered vulnerability discovery, dynamic validation, exploitability assessment, patch generation and human review. The company said a sample of 166 exploits contained at least 10 novel exploitation techniques.
By the time the research and manual review finished, 61 of the vulnerabilities had been patched upstream, 43 had staging patches, and 299 were being reported or discussed publicly, according to OpenAI. The company said it had sent hundreds of reports and patches for critical vulnerabilities and was working with maintainers and partners to submit the rest.
The project expands OpenAI’s Patch the Planet initiative and builds on its earlier cybersecurity-model research aimed at finding and fixing vulnerabilities in open-source software.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
