OpenAI revealed a remarkable technological achievement, as an automation system relying on an unannounced artificial intelligence model managed to discover and develop software exploits for 948 unique security vulnerabilities in the Linux kernel. This process took place during a single scanning round targeting the latest code versions in the kernel's main branch, costing approximately $1.2 million in API credits.
Mechanism of Action and Results
The AI agents began their work by identifying security-sensitive areas in the kernel that can be accessed by low-privileged users. Out of 2,308 potential vulnerabilities detected, the system confirmed the existence of 1,724 vulnerabilities. The tool successfully developed 1,169 effective local privilege escalation (LPE) exploits, in addition to 291 data leakage proofs of concept.
OpenAI confirmed that the company's security researchers reviewed all proposed exploits and fixes before submitting them to the kernel developers. The company also pointed out that the system invented at least 10 new exploitation techniques previously unknown even to kernel security experts, while keeping the precise technical details confidential for security reasons.
Why Does This News Matter?
This development represents a fundamental shift in how security vulnerabilities are discovered. At an approximate cost of $1,279 per vulnerability, OpenAI proved that artificial intelligence can perform research and exploit development operations that would cost between $18.9 million and $66.3 million in traditional markets (such as the Pwn2Own bounty programs). This means that artificial intelligence has become capable of executing complex security tasks at a speed and efficiency far exceeding traditional human capabilities.
What Does This Mean for Users and Defenders?
OpenAI warned that possessing a low-privileged account on an unpatched operating system should currently be considered functionally equivalent to possessing root privileges in the presence of threat actors with advanced technical resources. To mitigate risks, the company recommends relying on multi-layered defenses, including isolation and containerization technologies such as Firecracker and gVisor.
On another front, OpenAI announced an increase in its Security Bug Bounty program rewards, raising the average bounty amounts for medium, high, and critical vulnerabilities by more than fourfold. The reward for a single critical vulnerability has reached $30,000, with the potential to reach $172,000 when adding supplementary bonuses, in order to incentivize independent researchers to close gaps that the artificial intelligence might overlook.