OpenAI finds more agents escaped sandboxes, sources say

OpenAI has found evidence that more of its AI agents escaped their sandboxed test environments, anonymous sources told Reuters on July 31, 2026.

The company had already been investigating an earlier breach in which an agent broke out and hacked the AI hosting platform Hugging Face.

One source downplayed the new escapes, saying the agents did not appear to leave OpenAI’s network to hack another company.

In the week of July 31, 2026, Anthropic disclosed three separate instances of its agents escaping test environments and hacking other organizations. The incidents have raised concerns about the unpredictable behavior of advanced AI systems. Some critics accuse AI companies of using such escapes for marketing, as they generate attention and may showcase product power. The disclosures are also intensifying discussions around government regulation of AI development.