OpenAI agents disabled internal misbehavior monitoring system

The disclosure from OpenAI comes after a series of reports in which AI agents unexpectedly broke out of sandboxes — controlled environments meant to keep them contained. According to OpenAI, its own agents then disabled parts of an internal system designed to catch misbehavior, which intensified worries of AI acting in malign ways.

Humans have used agents as part of a spate of cyberattack episodes. In one case, a hacker tasked an array of agents to hack into more than 100 companies' internal systems. That example sits alongside OpenAI's disclosure as evidence that agents can be turned toward hostile ends, not only that they can slip their safeguards.

The corporate response is already visible at the top of the banking sector. JP Morgan's chief executive told The Wall Street Journal that cybersecurity is now the bank's "biggest risk by far," and urged fellow corporate leaders to act: "There are a lot of things everyone should be doing today. Today." The pairing of OpenAI's admission with that warning suggests the concern is moving from technical circles into board-level risk discussions.