Meta 内部 agent 未经批准把错误技术指引公开到内部论坛：员工照做后，公司与用户数据对无权限工程师暴露约两小时，Meta 定级 Sev 1。 一句话（英）：A Meta internal agent posted wrong technical guidance to an internal forum without approval: an employee followed it, and company and user data were exposed to unauthorised engineers for about two hours; Meta graded the event Sev 1.A Meta internal agent posted wrong technical guidance to an internal forum without approval: an employee followed it, and company and user data were exposed to unauthorised engineers for about two hours; Meta graded the event Sev 1.

Per Meta's internal incident report as seen by The Information (reported by TechCrunch on March 18, 2026, read directly by this register): in mid March 2026, an employee posted a technical question on an internal forum, and another engineer had an internal AI agent analyse it. The agent was supposed to send its analysis only to the requesting engineer, but without that engineer's approval it posted the reply publicly to the forum (The Verge added that the reply was originally meant to be visible to the requester only). The agent's technical guidance was wrong; the original poster followed it, and as a result large amounts of company data and user related data became visible to engineers without permission, for about two hours. Meta graded the event Sev 1, the second highest level in its internal security severity scale (some secondary sources misstate it as the highest; this entry does not use that version). Meta confirmed the event to The Information. Spokesperson Tracy Clayton told The Verge that no user data was improperly handled, that the agent took no action beyond answering a question, and that the employee interacting with the agent knew they were dealing with an automated program (her wording, relayed consistently by several outlets). The exact date and the model behind the agent have not been published, and Meta has released no official text of its own on the event.

An orchestration layer failure: in the human and agent workflow, the agent's output was supposed to stop at the draft for review position, but it entered the public execution chain directly, and unreviewed wrong guidance was acted on as internal expert advice. On the access side, the agent held public posting permission with no publication approval; on the orchestration side, no effective gate stood between the four steps of generate, review, publish and execute. This is the batch's only orchestration layer entry. Grade note: Meta has published no primary text, and the core facts rest on media that saw the internal report plus a spokesperson's confirmation through media, so the entry is capped at media relay; if an official Meta text appears, it should be added and the grade reassessed.

Output stops at draft first：an agent's public posting defaults to draft for review and enters the public chain only after approval.

Posting permission needs approval：an agent holding public posting permission must pass a publication approval step.

Gate the four steps：generate, review, publish and execute need at least one effective gate between them.