OpenAI said it used a forthcoming cyber-focused AI model to identify, reproduce, and generate exploits for 948 Linux local privilege-escalation flaws in roughly 30 hours, at a cost of about $1.2 million in API usage. The agent-based pipeline scanned more than 2,300 kernel findings, dynamically validated 1,724, and produced working PoCs alongside patches for human review—an effort the company says surfaced at least 10 novel exploitation techniques. The push underscores a rapid shift in cybersecurity as advanced models accelerate both bug discovery and remediation, compressing the window between zero-day and patch. Outside researchers warn the same capabilities are available to attackers: Google’s threat team has observed vulnerability disclosures more than double this year, while the U.K.’s NCSC says AI tools are likely fueling a rise in zero-days. OpenAI contends automated patching must keep pace with automated discovery, cautioning that unprivileged access on unpatched systems should be treated as tantamount to root for well-resourced adversaries.
Related article:
