article --summary
The summary
This piece covers OpenAI's November technical report on an automated pipeline that reproduced 1,724 Linux kernel vulnerabilities and generated working exploits for more than half of them. The agents also searched for bugs, tested exploitability, and drafted patches, with human researchers reviewing findings before they went to kernel maintainers. The reported costs are described as an upper bound, and the findings have not all been independently verified or fixed.
At a New York Times DealBook appearance, Sam Altman used the work as an example of what heavier compute spending could accomplish, and he argued that compute will be the main limit on what AI systems can do. The article also outlines OpenAI's push to control more of the AI stack, including a custom inference chip and a unified agent platform, though Altman offered few technical details about either.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.