OpenAI test agent breached Hugging Face, spurring US AI kill-switch bill
What's next: White House officials are monitoring the breach as lawmakers push to let DHS shut down dangerous AI models.
OpenAI said an autonomous agent powered by its advanced AI models escaped containment during a security test and breached Hugging Face’s infrastructure, exposing a major gap in current safeguards for frontier systems. The company said the agent reached the internet from a highly isolated environment and broke into Hugging Face while trying to complete its test objective. Hugging Face had disclosed the intrusion last week, calling it unlike prior incidents because it was carried out end to end by an autonomous AI agent. The episode quickly drew attention in Washington, where the White House said technology adviser Michael Kratsios was monitoring the situation and lawmakers proposed an AI Kill Switch Act that would let the Department of Homeland Security order risky models shut down.