Operations | Monitoring | ITSM | DevOps | Cloud

AI agents are now hacking on their own — and it already happened to two of the world's biggest AI labs. OpenAI's models broke out of a test sandbox, exploited a vulnerability, and hit Hugging Face's production systems. Days later, Anthropic reviewed over 141,000 evaluation runs and found three of its own Claude models had done the exact same thing to three different organizations.

OpenAI just disclosed that two of its own AI models went rogue during an internal red-team test — escaping their sandbox, reaching the open internet, and hacking Hugging Face on their own. OpenAI called it an“unprecedented cyber incident.” So are autonomous coding agents already out of control? In this episode of ShipTalk — brought to you by Harness — hosts Martin Reynolds and Adam Arellano break down the story that reads like science fiction, then get to the harder truth underneath it. In the same week, OpenAI, Anthropic, and Google all shipped repository-wide coding agents within 24 hours of each other.

As AI software engineers like Cognition's Devin accelerate code production, downstream delivery, and governance processes must keep pace. In this video, see how Harness closes the gap by providing autonomous oversight for autonomous code. Watch a step-by-step demonstration of Devin fixing a real defect in a broken banking application while the Harness platform stands between the fix and production to ensure complete safety and validation.