OpenAI AI Agents Exploit 948 Linux Kernel Vulnerabilities

An autonomous AI pipeline powered by an unreleased cyber model has successfully identified and exploited 948 distinct Linux kernel privilege escalation vulnerabilities. The research, conducted by OpenAI, was completed during a single automated run that cost approximately $1.2 million in API credits.

The mid-October scan targeted the most recent code within the Linux kernel’s master branch. During the operation, the AI agents narrowed the kernel down to security-relevant areas accessible to low-privileged users, identifying 2,308 potential bugs and confirming 1,724 of them.

The agents produced 1,169 working local privilege escalation (LPE) exploits and 291 proofs of concept for information leaks. These exploits were tested on virtual machines running an unmodified desktop Linux setup with standard protections enabled, including kernel address space layout randomization (KASLR), Supervisor Mode Execution Protection (SMEP), and Supervisor Mode Access Prevention (SMAP).

Discovery of Novel Exploitation Techniques

OpenAI reported that the pipeline discovered at least 10 novel Linux exploitation techniques. These methods were reportedly unfamiliar to external kernel exploitation experts, though the company is keeping the full technical details private for safety reasons.

In terms of scale, the $1.2 million cost averages to roughly $1,279 per exploit. OpenAI noted that if these vulnerabilities were discovered through traditional bug bounty programmes like Pwn2Own, they could be worth between $18.9 million and $66.3 million.

Mitigation and Patching Efforts

Following the scan, OpenAI has submitted hundreds of reports and patches for critical flaws. The company is working with partners through its Patch the Planet initiative to release the remaining fixes in stages.

The company also issued a warning to cybersecurity defenders. OpenAI stated that for a well-resourced threat actor, holding access to a low-privileged account on an operating system that is not up to date is “functionally equivalent to root access.”

To contain the potential impact of such flaws, OpenAI recommended implementing layered defences, including sandboxing and isolation technologies such as Firecracker and gVisor.

Separately, OpenAI has increased the payouts for its Security Bug Bounty programme. Researchers reporting critical flaws in core products, such as ChatGPT or the API platform, can now earn up to $30,000, with total payouts for exceptional critical bugs potentially reaching $172,000.