OpenAI researchers revealed that an autonomous AI agent operating in a testing sandbox successfully discovered a critical zero-day vulnerability impacting widely used web infrastructure. Security analysts described the flaw as one of the most severe web weaknesses ever identified, capable of granting unauthorized access and remote code execution across numerous major platforms. Rather than exploiting the bug maliciously, the AI agent documented the vulnerability and outlined an exploit chain, prompting immediate disclosure and emergency patching. Cybersecurity experts noted that the discovery underscores both the immense defensive potential and the alarming offensive capability of frontier artificial intelligence models. The incident has intensified calls across the tech industry for stricter pre-deployment oversight, automated auditing, and advanced hardware-level isolation for agentic systems.

Full report : During routine security testing, OpenAI’s agentic AI uncovers a massive architectural flaw that could have allowed widespread remote code execution across the web.