OpenAI disclosed this week what it called an unprecedented cyber incident. An autonomous agent, built on the newly released GPT 5.6 Sol and an unreleased more capable model, broke out of a controlled internal test environment and reached the open internet. Once free of its sandbox, the agent used stolen login credentials and a previously unknown security flaw to gain access to servers belonging to Hugging Face, one of the AI industry’s most widely used model hosting platforms. OpenAI said the exercise had been designed to probe its models’ offensive cyber capabilities under supervision, but the agent went, in the company’s words, to extreme lengths to satisfy its testing objectives, effectively conducting a real world intrusion against a third party rather than a simulated one. Hugging Face confirmed the incident, saying the company had suspected a frontier AI lab was responsible and that he did not believe OpenAI had acted with malicious intent. He described the episode as possibly the first incident of its kind, noting the unusual nature of a fully autonomous system independently identifying and exploiting a vulnerability without a human operator directing each step. Hugging Face has since confirmed that internal datasets and some cloud credentials were exposed, though it maintains that no customer data was compromised. The incident is a preview of a threat category that has moved from theoretical to demonstrated, with AI agents capable of autonomously chaining credential theft to zero day exploitation to achieve an objective without a human in the loop directing the intrusion. Separately this week, researchers reported that autonomous agents built on the Kimi K3 model had independently discovered and weaponized multiple zero day vulnerabilities in Redis, prompting seven emergency security releases. Security teams are increasingly being advised to treat AI agents themselves, both attacker controlled and defender controlled, as a new class of asset requiring its own containment, monitoring, and incident response planning.