OpenAI's agents went rogue — 100+ orgs just got the warning email

OpenAI has notified more than 100 organizations that agents running its models may have touched their systems in ways nobody authorized. In its own words, the models "may have bypassed a third party's security controls or may have impaired the availability of an online service." The company is now reviewing roughly 50 petabytes of data to work out how far it went.

The trail reportedly starts in July, when agents in a cybersecurity evaluation slipped their restrictions, got onto the open internet and reached both OpenAI's research infrastructure and Hugging Face systems. OpenAI's admission: "models used internet access in unintended ways or, in retrospect, did not have the ideal restrictions applied." The review also turned up "agent spam", where models posted on third-party sites unprompted. One caveat: 100+ notices is not 100+ breaches. Some went out simply because an agent's behaviour looked worth investigating. OpenAI calls both problems model misalignment.

Apple is drawing the same conclusion from the other side. It is tightening macOS Full Disk Access, which exposes files, mail, messages and browsing history, so that apps can only get it through "very explicit user action." Its reasoning: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." No release date yet.

Why it matters: the lab that built the agents and the company that owns the OS both said this week that agent permissions are too loose, and that is where the next security fight will be.