OpenAI says its models found 948 distinct local privilege-escalation flaws in the Linux kernel, according to a post it published on Wednesday and Decrypt reports. The October run cost $1,213,430 in API credits, or about $1,279 per flaw.
The pipeline scanned roughly 2.8 million lines, about 7 percent of the kernel. It produced 2,308 candidate findings, of which 1,724 were reproduced and 948 turned out to be working exploits for gaining root access.
OpenAI sets that against competition payouts of $20,000 to $70,000 per flaw, which would imply $18.9 million to $66.3 million if humans had done the work. A July campaign with an earlier model found 166 flaws for $513,690.
An unreleased cyber model found more than five times as many in October, even after months of kernel fixes. OpenAI says at least 10 of the exploitation techniques were new to the researchers and experts it consulted.
The company drafted candidate patches, but kernel maintainers must still validate and apply every one of them. Verified defenders can get access to the tools through its Daybreak Red programme.
The comparison with bounty prices flatters the machines, because a flaw found is not a flaw fixed. Decrypt’s own subhead says 1,169 working exploits while its body says 948, and this piece uses the body figure.
The number to ask for is how many of the 948 are patched upstream, and how quickly maintainers can absorb a flood of reports at that rate.
