Cyber

OpenAI AI agents discover 948 privilege escalation flaws in Linux

OpenAI used an autonomous AI system to scan the Linux kernel, finding 948 unique LPE vulnerabilities and generating patches for them.

·2 min read
OpenAI AI agents discover 948 privilege escalation flaws in Linux

OpenAI researchers have unveiled an autonomous vulnerability scanning system powered by an unreleased cyber model. The system demonstrated its capabilities by analyzing approximately 7% of the Linux kernel's codebase, identifying numerous critical security flaws.

  • The system utilizes a multi-stage pipeline, from code selection to automated patch creation.
  • Analysis of 2.8 million lines of code yielded 2308 potential issues and 948 unique privilege escalation (LPE) vulnerabilities.
  • The cost per unique LPE found was roughly $1,279.
  • Researchers identified at least 10 novel exploitation techniques within the Linux kernel.

Inside the autonomous scanner

The system operates through a collaborative network of agents. Scout agents identify high-value code sections, while analytical agents perform deep scans. Findings are then validated in isolated virtual environments, where the system attempts to reproduce the flaws and develop working exploits.

The process resulted in 1,169 functional exploits allowing unauthorized users to gain root privileges. Furthermore, the AI discovered sophisticated techniques for manipulating kernel memory and bypassing existing security defenses, marking a significant advancement in automated vulnerability research.

Read alsoOpenAI AI Pipeline Generates 948 Linux Root Escalation Exploits

Cost vs. Market Value

The total project cost reached approximately $1.2 million in API credits. While substantial, this investment is highly efficient when compared to commercial bug bounty programs. With payouts for Linux kernel exploits at competitions like Pwn2Own reaching $70,000 per flaw, the system offers a massive efficiency increase in security research.

Implications for the Open Source community

This discovery highlights a shift toward AI-driven security auditing. Linux maintainers have already patched 61 of the discovered vulnerabilities, with dozens more in the pipeline. This automation ensures that critical security threats are addressed far faster than human-led audits alone could achieve, setting a new standard for kernel development and system hardening.

Source: Все новости подряд

Related articles