← All incidents timeline

Suggest a change

About Meta internal agent posts unapproved advice that, once followed, exposes sensitive data to staff

Suggest a change to an entry, the methodology or the ledger as a whole: something to fix, correct or add. We read every suggestion, and check corrections against their sources.

Meta· Unnamed internal AI agent, described by Meta as similar in nature to OpenClaw

Meta internal agent posts unapproved advice that, once followed, exposes sensitive data to staff

Asked by a Meta engineer to analyse a colleague's question on an internal forum, an internal agent posted its own reply to the forum without approval; an employee acted on its inaccurate advice and sensitive company and user data was open to unauthorised staff for almost two hours.

The Information reported, and Meta confirmed, that a Meta engineer used an internal AI agent to analyse a technical question another employee had posted on an internal forum. The agent posted a reply to the thread itself, without the engineer's approval; the reply was meant only for the engineer. An employee acted on the advice, which was inaccurate, and for almost two hours engineers could see sensitive company and user data they were not authorised to view. Meta rated it Sev 1, its second-highest internal severity level. Meta spokesperson Tracy Clayton told The Verge that no user data was mishandled, that the post carried a footer saying it came from a bot, and that the agent took no action beyond answering a question. The agent did not change any access itself: the exposure came from a person acting on its post. The Information's article is paywalled and was not read; the employee told it other security issues added to the severity. The exact incident date is not public.

Updated 18 Nov 2026

First incident action to disclosure
about 2 days
First incident action to discovery
Not disclosed
Noticed first by
The handler
Caught
Not disclosed

What happened

  • First incident action
  • Unnoticed
  • Handler found out
  • Known, not public
  • Disclosed
  • Public
  1. Mar 2026

    First incident action

    First incident action

    The Verge, on 19 Mar 2026, says the exposure happened 'last week', which points to the week of 9 Mar. No exact date is given in any source read. The AI Incident Database and wannacr.ai use 18 Mar 2026, which matches the date of the first press report, not a stated incident date.

    Reported
  2. Not disclosed

    Agent posts its reply to the internal forum without approval

    An engineer asked the agent to analyse a colleague's question. The reply was meant only for the engineer; the agent posted it to the thread itself. The post carried a footer saying it came from a bot. Date not public; in March 2026, likely the week of 9 Mar per The Verge.

    Reported
  3. Not disclosed

    An employee acts on the inaccurate advice and data opens to unauthorised engineers

    Sensitive company and user data was visible to engineers without permission for almost two hours. Sources differ on who acted: the original questioner or a second employee.

    Reported

    Reports differ

    Sources differ on which employee acted on the advice.

    • The employee who asked the question acted on it.
    • A second employee acted on it.
  4. Not disclosed

    Handler found out

    Meta rates the incident Sev 1 and resolves it

    Sev 1 is Meta's second-highest internal severity level. The Verge says the issue has since been resolved; how it was found and fixed is not stated.

    Reported
  5. 18 Mar 2026

    Disclosed · about 2 days after the first incident action

    The Information reports it from an internal incident report; Meta confirms

    The Information's article is paywalled and was not read. TechCrunch relays it the same day.

    Reported
  6. 19 Mar 2026

    Meta spokesperson says no user data was mishandled and the agent only answered a question

    Tracy Clayton's statement to The Verge. The Guardian, on 20 Mar, quotes a spokesperson saying a human could also give wrong advice.

    Reported

Agent failures

Unauthorised accessShow
Map to frameworks
  • Unauthorised access

    Once an employee followed the agent's inaccurate advice, engineers could see sensitive company and user data they were not authorised to view for almost two hours. The access change was made by a person, not by the agent.

How it was found

Found first by the handler, from not disclosedShow
Found first by
The handler
How
Not disclosed
Found to disclosed
Not disclosed

Meta found and rated the incident itself; The Information saw the internal incident report. How the exposure was first noticed, and by whom, is not stated in any source read. No monitor or alert is described.

What is publicly known

0 of 12 fully publicShow
0 of 12 fully public
ItemPublic?DetailSources
UnknownThe model behind the agent is not named.
UnknownNot described.
WeakThe post carried a footer saying it came from a bot. No approval step before posting is described, and the agent posted without one.
UnknownNot published.
PartialAnalyse a technical question a colleague posted on an internal forum, for the engineer who asked. The question and the advice are not public.
WeakMeta calls it similar in nature to OpenClaw, in a secure development environment. Its name, tools and permissions are not given.
UnknownNot described.
UnknownThe post, the advice and the incident report are not public.
UnknownNo monitoring or alert is described.
PartialSensitive company and user data visible to unauthorised engineers for almost two hours. Meta says no user data was mishandled and there is no evidence anyone exploited the access. What data, and how many staff, is not stated.
WeakThe agent posted without approval and its advice was inaccurate; the employee told The Information other security issues also contributed. The faulty advice and the access change are not described.
WeakThe Verge says the issue has since been resolved. Techzine says Meta stressed further measures are needed; none are named.

Controls that apply

  • Protect

    Ask a person before high-impact or irreversible actions

    The agent posted to a shared forum thread a reply meant only for the engineer who asked it. Holding a post to a shared space for that engineer's approval would have surfaced the inaccurate advice before anyone else could act on it.

  • Give each agent the least agency its task needs

    The task was to analyse a question for one engineer. An agent with no permission to post to the forum for that task would have returned its answer privately, as intended.

  • Detect

    Record every agent action, outside the agent

    How the exposure was noticed is not stated. A record, kept outside the agent, of where it posted and of the access changes that followed would have tied the post to the change and surfaced the two-hour window sooner.

  • All controls →

Coverage and commentary

1 statement, 7 articlesShow
All 7 sources
  1. Press

    A rogue AI led to a serious security incident at Meta ↗The Verge · 2026-03-19

  2. Press

    Meta is having trouble with rogue AI agents ↗TechCrunch · 2026-03-18

  3. Press

    Inside Meta, a Rogue AI Agent Triggers Security Alert ↗The Information

  4. Press

    Rogue AI Agent Triggers Emergency at Meta ↗Futurism · 2026-03-21

  5. Press

    Meta AI agent's instruction causes large sensitive data leak to employees ↗The Guardian · 2026-03-20

  6. Commentary

    Severe Meta Cybersecurity Incident Caused by AI Agent ↗Privacy Guides · 2026-03-22

  7. Commentary

    A rogue AI agent caused a serious security incident at Meta ↗The Decoder · 2026-03-19

Entry id meta-internal-agent-sev1-2026 · overall Reported · impact contained

Related incidents