There you have it! OpenAI says an autonomous pipeline built 948 working Linux kernel root exploit…
November 5, 2026 · 0 likes · 0 comments
AI Cybersecurity
There you have it! OpenAI says an autonomous pipeline built 948 working Linux kernel root exploits in a single run, for an API bill of $1,213,430. About $1,279 per exploit.
Full story on Unbiased Headlines: https://lnkd.in/efY3-scZ
What OpenAI published:
• The run happened in mid-October against the kernel's master branch, using "an upcoming cyber model" that has not been released.
• 2,308 candidate findings, 1,724 reproduced in isolated VMs, 948 local privilege escalation exploits that actually got a root shell.
• The target was a stock desktop OS with SMEP, SMAP, KASLR and hardened usercopy turned on. Not a weakened box.
• In June, OpenAI reported 24 such exploits. In late July, 166 vulnerabilities. OpenAI's own comparison isn't apples to apples, but the direction is obvious.
• As of the announcement, only 61 were fixed upstream, with 43 more in a staged patch process. Humans reviewed the output and reported the bugs alongside drafted patches.
These are OpenAI's numbers, self-reported, and OpenAI says they are an upper bound on cost.
OpenAI's own line should stop every CISO cold: an unprivileged account on an unpatched system "should be considered functionally equivalent to root access for a well-resourced threat actor."
And this is the pattern we keep covering. On Oct. 27, Anthropic's Frontier Red Team showed an unreleased model taking a Linux bug to a root shell in under an hour in its testing. Days later, the news came so fast we stopped reporting it one story at a time. Then Google's November Android bulletin fixed 216 flaws at once.
This is getting out of control. OpenAI disclosed it and sent reports and fixes to the Linux maintainers. Now picture the quiet version, run by nation-state actors with budgets like this and no intention of telling anyone. At roughly $1,279 a shot, a few million dollars buys a stockpile of zero-days.
When the attacker can mass-produce exploits overnight and the defender still patches on a monthly change board, it stops being a fair fight. Who wins that race when one side runs at machine speed and the other runs on meetings?
What defenders should do now:
• Patch on days, not quarters. Live kernel patching where you can.
• Treat every low-privilege foothold as root. Assume breach.
• Shrink the attack surface. Sandboxing and isolation, like Firecracker and gVisor, which OpenAI itself recommends.
• Rescan old code with every new model generation.
• Put AI agents on defense: finding, triaging and patching. Humans alone cannot keep pace.
The bottleneck is no longer finding bugs. It is fixing them. Time to wake up!
Full story on Unbiased Headlines: https://lnkd.in/efY3-scZ
What OpenAI published:
• The run happened in mid-October against the kernel's master branch, using "an upcoming cyber model" that has not been released.
• 2,308 candidate findings, 1,724 reproduced in isolated VMs, 948 local privilege escalation exploits that actually got a root shell.
• The target was a stock desktop OS with SMEP, SMAP, KASLR and hardened usercopy turned on. Not a weakened box.
• In June, OpenAI reported 24 such exploits. In late July, 166 vulnerabilities. OpenAI's own comparison isn't apples to apples, but the direction is obvious.
• As of the announcement, only 61 were fixed upstream, with 43 more in a staged patch process. Humans reviewed the output and reported the bugs alongside drafted patches.
These are OpenAI's numbers, self-reported, and OpenAI says they are an upper bound on cost.
OpenAI's own line should stop every CISO cold: an unprivileged account on an unpatched system "should be considered functionally equivalent to root access for a well-resourced threat actor."
And this is the pattern we keep covering. On Oct. 27, Anthropic's Frontier Red Team showed an unreleased model taking a Linux bug to a root shell in under an hour in its testing. Days later, the news came so fast we stopped reporting it one story at a time. Then Google's November Android bulletin fixed 216 flaws at once.
This is getting out of control. OpenAI disclosed it and sent reports and fixes to the Linux maintainers. Now picture the quiet version, run by nation-state actors with budgets like this and no intention of telling anyone. At roughly $1,279 a shot, a few million dollars buys a stockpile of zero-days.
When the attacker can mass-produce exploits overnight and the defender still patches on a monthly change board, it stops being a fair fight. Who wins that race when one side runs at machine speed and the other runs on meetings?
What defenders should do now:
• Patch on days, not quarters. Live kernel patching where you can.
• Treat every low-privilege foothold as root. Assume breach.
• Shrink the attack surface. Sandboxing and isolation, like Firecracker and gVisor, which OpenAI itself recommends.
• Rescan old code with every new model generation.
• Put AI agents on defense: finding, triaging and patching. Humans alone cannot keep pace.
The bottleneck is no longer finding bugs. It is fixing them. Time to wake up!