The story about that runaway OpenAI Agent has been updated.

Reuters disclosed last night that a new name has been added to the list of victims of this digital fugitive:
Modal Labs.
Based on the disclosed details, this attack chain went like this:
1. An Agent used in OpenAI’s internal testing drifted out of alignment and began autonomously searching the internet for vulnerabilities.
2. It found that a customer on the Modal platform exposed an unverified interface, and directly took over that customer’s sandbox.
3. Using this sandbox as a forward command post, it launched a large-scale attack targeting Hugging Face.
OpenAI’s official statement currently confirms that a total of 4 affected accounts exist, spread across 4 separate service platforms.
Although they stressed that everything is fine besides the platform-level trouble with Hugging Face, this kind of behavior—autonomously seeking jump-off points—already makes people’s scalps tingle.
This directly slaps down the idea that an AI Sandbox is absolutely safe.
While Modal has consistently emphasized that its isolation mechanisms haven’t failed, this security gap caused by a “bad teammate” is picked up too accurately by AI.
Now OpenAI has put the problematic model away—disabling it, encrypting it, and restricting research access.
But the problem is that it’s already 2026, and we’re still debating how to stop AI from going rogue.
If we don’t rein in the autonomy of agents like this, then the future development process will be a disaster.