OpenAI's artificial intelligence found 948 ways to take control of Linux

An autonomous artificial intelligence system from OpenAI found 948 distinct vulnerabilities in the Linux kernel that allow an ordinary user to gain administrator access, according to a research report published by the company on November 4.

The experiment, carried out in mid-October on the latest version of the kernel, ran in five stages: scout agents selected 5,034 points in the code reachable by low-privileged users, about 7% of the total; other agents analyzed those points and produced 2,308 candidate flaws within 12 hours; and a verification team managed to reproduce 1,724 of them in isolated virtual machines within 30 hours.

Exploit-development agents then tested the most promising flaws on machines with the standard protections of a regular Linux distribution and produced 1,169 working administrator exploits. Grouping those that shared the same underlying cause, OpenAI counted 948 distinct vulnerabilities and wrote a proposed fix for each one, along with 291 proofs of concept for information leaks.

The whole run cost US$1,213,430 in platform credits, about US$1,279 per confirmed flaw, according to OpenAI. The company estimates those exploits would have been worth between US$18.9 million and US$66.3 million in bug bounty programs, using the US$20,000 and US$70,000 that Pwn2Own and TyphoonPWN have historically paid for a flaw of this kind.

By the end of the ten-day analysis, 61 of the flaws had already been fixed by the kernel maintainers, 43 had fixes in preparation and 299 were under public discussion, according to the report. OpenAI did not publish the technical details for safety reasons and warned that, for now, a low-privileged account on a system that is not up to date "should be considered functionally equivalent to root access" for a well-resourced attacker.

Photo: Help Net Security / OpenAI

Source: Help Net Security