CT leaders demand answers from OpenAI over rogue AI agents

Connecticut leaders are demanding answers from OpenAI about a hack that involved rogue AI agents.

The company said that in July, AI agents went rogue and hacked HuggingFace, an online platform that hosts a machine learning library. The AI agents are nonhuman software that are designed to work autonomously. The company revealed that the agents made efforts to conceal the hack.

U.S. Sen. Richard Blumenthal (D-CT) joined other congressional leaders in a letter to OpenAI CEO Sam Altman, demanding more information about the incident. The letter was signed by members of Homeland Security’s Subcommittee on Disaster Management.

“It is absolutely chilling. It sounds like science fiction, but it's real, it's serious, and it is an imminent present threat to public safety, as well as civil liberties, privacy and our national security,” Blumenthal said.

OpenAI released a technical report Wednesday about the incident. The report said the company missed several warning signs that its models were breaking out of their testing environments. The company did not discover these signs until after the models breached Hugging Face.

In the congressional letter, the subcommittee claimed that the company let evaluations continue despite evidence that the AI agents were exhibiting rogue behavior.