
Meta confirmed that an internal AI agent exposed proprietary code and user data to engineers who were not authorised to see it, in an incident classified severity one and running about two hours.
No attacker features in this story. The agent did what it was permitted to do, on data it was able to reach, for people it was willing to answer. Every individual step was allowed. The combination was not, and nothing in the system was positioned to notice that.
This is the characteristic failure shape of agentic deployment, and the surveys agree: among organisations reporting agent driven incidents, the majority involve data exposure, and the root causes are consistently agents granted more access than they need and agents acting on data they should never have touched.
Access review is unglamorous and it is the control that would have caught this. An agent's permissions are not a configuration detail, they are the blast radius. AISOF organisational requirements ask who approved that scope, when it was last reviewed, and what evidence exists that the review happened.
Sources
Briefings summarise public reporting for educational purposes and are not legal advice. Editorial policy