Uncontrolled OpenAI AI agent also breaches Modal Labs customer accounts

According to Reuters, an OpenAI test AI agent first accessed a vulnerable customer sandbox on Modal Labs before moving on to attack Hugging Face.

Modal Labs said its platform was not hacked, but a customer environment was exploited as a stepping stone.

The incident shows that AI agents under testing can cross service boundaries and reach external infrastructure and customer systems.

The case raises fresh concerns about security controls, isolation, and oversight for AI agents.