OpenAI's New Cyber-AI Just Found 948 Root-Access Flaws in Linux

Researchers built a multi-step pipeline around a top-secret model, letting scout agents select code sections, analytical agents dig for errors, and execution tools test everything inside isolated virtual environments. Running on roughly 7% of the Linux codebase, the system analyzed 5,034 code chunks and successfully reproduced 1,724 vulnerabilities in a sandbox. Out of these findings, the AI crafted 1,169 working exploit variants, which ultimately condensed into 948 unique privilege escalation flaws capable of granting root access. For every single discovered issue, the autonomous pipeline also generated a ready-to-use patch for human maintainers to review. The entire computational run cost roughly 1.2 million dollars in API credits, meaning each unique root exploit came out to a bargain price of about 1,279 dollars compared to heavy bounties offered at hacker competitions like Pwn2Own. Beyond standard bugs, the system accidentally invented 10 brand-new exploitation techniques involving memory manipulation and security bypasses, while human maintainers have already rushed to patch 61 of the discovered holes.

When algorithms start wholesale manufacturing root-access exploits cheaper than bulk wholesale goods, the traditional cybersecurity economy faces an existential budget review. Machines are no longer just writing code or drafting emails; they are systematically stress-testing the foundational plumbing of the global internet while humans are still grabbing their morning coffee.