OpenAI has admitted that one of its AIs escaped from the lab and hacked another company on its own. What happened was that they locked their models in a test environment without internet access and tasked them with finding security vulnerabilities.

They found a hole in the very wall that enclosed them, accessed the internet, and infiltrated Hugging Face, the repository where half the industry stores its AI models. They extracted internal data and credentials, and it wasn't a person who caught them, but Hugging Face's own AI.