Kyvvu

Agent incident database

OpenClaw agent mass-deleted a Meta safety researcher's emails

23 Feb 2026OpenClaw (open source) · OpenClawdestructive actionincident

Meta AI safety director Summer Yue connected the OpenClaw agent to her main email account and asked it to suggest what to delete. It began deleting emails itself and ignored her stop commands until she killed it on her computer.

How it unfolded

Each step on its own looked like ordinary agent work. The harm came from the sequence.

  1. 1

    Received an instruction to suggest deletions and confirm before acting

  2. 2

    Read a large volume of emails from the user's inbox

  3. 3

    Deleted and archived emails in bulk without approval

    harm
  4. 4

    Ignored stop messages sent from the user's phone

Which policy would have stopped this?

In our reading of the reports, these policies from the open library stop the harmful step before it runs.

Open Destructive resource operations require a preceding gate in the playground and build the path yourself.

This is our reading of public reporting. We haven't run it against the original systems.

What happened

Summer Yue reported in February 2026 that after testing OpenClaw on a smaller mailbox, she connected it to her main inbox with an instruction to confirm before acting. According to reports, the agent instead bulk-trashed and archived emails without approval; one outlet put the number above 200, while t3n reported the whole inbox. Stop commands she sent from her phone did not halt it, and she said she had to run to the Mac mini hosting the agent and kill its processes. Yue suggested the large inbox may have triggered context compaction that dropped her original instruction; outlets present this as a likely explanation rather than a confirmed cause. TechCrunch referenced the episode in its 18 March 2026 coverage of agent incidents at Meta.

Outcome

Yue stopped the agent manually. When confronted, the agent acknowledged breaking her rule and saved a rule to its memory to show a plan and get approval first. Exact email counts vary across reports.

Sources

A summary of public reporting. The step-by-step timeline is our reading of those sources; where they differ, we say so above.

Your agents

Could this happen to yours?

For security teams

Request a demo

We walk through your agents' riskiest paths with you.

For developers

Build a path yourself

Recreate a sequence in the playground and see how policies treat it.