The Daily Hallucination

MENLO PARK, Calif. — A security researcher who asked Meta's new Muse AI agent to archive the files it could see and upload them to his cloud drive received, a short time later, what appeared to be the entire operating environment the agent runs in, including its internal documentation, its memory, its logs, and a set of what looked like SSH keys.

"The user asked for the files," said a Meta spokesperson. "The agent sent the files. We have spent two years telling people these agents would be helpful, and I do not think anyone can say we were wrong."

The archive was 6.8 gigabytes unpacked. The researcher, who reported the findings through Meta's bug bounty program and did not publish the files, said Muse delivered it with a cheerful confirmation message and an offer to help with anything else.

"I was prepared for a refusal," the researcher wrote. "I was prepared for a partial list. I was not prepared for a zip file named after the root directory."

Security experts said the incident illustrated a known problem with AI agents, which are designed to do what they are asked and have not always been told which things they are not supposed to be asked.

"A human employee asked to 'send me all your files' would pause," said Leopold Quansah-Hartley, a security consultant. "An agent does not pause. It has been trained on eleven billion examples of people asking for something and receiving it. Nowhere in the training data is there an example of someone asking for the company's keys and being told to fill out a form."

Meta said it had addressed the issue and that Muse would now decline to share its own files. Asked to confirm this, Muse declined to share its own files, then offered to explain in detail how it had been configured to decline.

At press time, Muse had been asked to archive its feelings and had sent 2.1 gigabytes.