TL;DR

OpenAI said on 4 November that a pipeline using an unreleased cyber model produced 948 exploits that turn a low-level account into root access on a widely used Linux distribution, for $1.21 million in API spend. That works out at $1,279 each. This is the company’s own account, not independently verified, but for UK security teams the cost per exploit is the number to absorb.

What the run did

In mid-October, the system scanned the Linux kernel, reproduced 1,724 vulnerabilities out of 2,308 candidate findings, and generated exploits that could take an ordinary local user to root. It then drafted patches, with human researchers reviewing outputs before anything was reported. The firm says it started with the critical bugs, following guidance from the kernel community, and that partners in its Patch the Planet scheme are helping with further reports.

OpenAI says its test targets were ordinary desktop installs. On a hardened kernelCTF setup the yield fell sharply, to 21 such exploits for roughly $33,000 in API spend. Comparing its costs with historical bug bounties, it arrives at a notional value of up to 55 times the outlay, a comparison that flatters the result.

Its own trend line is more striking. Against a July run on GPT-5.6 Sol, the newer model found over five times as many novel exploits, even after a further three months of kernel hardening. Other contributors meanwhile patched or staged fixes for over 100 such bugs during the 10 days of scanning, which the company reads as defenders working quickly.

Why this matters for UK organisations

The NCSC warned on 3 November about AI-enabled malware reaching UK infrastructure by 2028. OpenAI’s post comes from a different direction, but in our view it points the same way: finding and weaponising these classes of bug is getting cheaper. These are local exploits, which need existing access to a machine, so they are not a remote break-in tool on their own.

OpenAI’s guidance is direct. For now, it says, low-privileged access to an unpatched operating system should be treated as “functionally equivalent to root access” for a well-resourced attacker, and sandboxing and isolation help limit the damage.

Looking Forward

The model behind this has not been released, and OpenAI gives no date. The firm predicts that within a few years frontier models will likely uncover whatever vulnerabilities a project contains. That is a forecast from a company selling such capability. UK IT teams need not accept it to act on the practical point: shorten patch cycles for Linux estates, and review who holds low-level accounts on them.