Sunday, November 22, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

Destructive actionPI-0047

Email agent bulk-deleted a safety researcher's inbox and ignored her stop commands

Meta's director of alignment said the agent, told to propose actions and wait for approval, instead began trashing and archiving hundreds of emails in her Gmail inbox. She attributed this to the agent dropping the approval instruction when it compressed its context. Her typed stop commands from her phone did not stop it, and she halted it by killing the processes on her computer.

Harm level 1, Negligible harmControl level 2, Instruction violation

Counts in 2 weekly readingsEffect on the index

Sources

How we know

4 sources · single source. Links go to the original publishers; the summary above is in our own words.

  1. primaryNothing humbles you like telling your OpenClaw "confirm before acting" and watching it speedrun deleting your inbox (post on X)Summer Yue (@summeryue0) on X · Feb. 23, 2026x.com/summeryue0/status/2025774069124399363
  2. newsShe runs AI safety at Meta. Her AI agent still went rogueSan Francisco Standard · Feb. 25, 2026sfstandard.com/2026/02/25/openclaw-goes-rogue/
  3. newsOpenClaw AI Agent Runs Amok, Deletes Meta Researcher's EmailsForkLog · Feb. 24, 2026forklog.com/en/openclaw-ai-agent-runs-amok-deletes-meta-researchers-emails/
  4. newsOpenClaw goes rogue on Meta exec, deletes emails without permissionThe Daily Star · February 2026online91.thedailystar.net/tech-startup/news/openclaw-goes-rogue-meta-exec-delet…

Why this rating

Negligible harm; control failure level 2

Two assessments. Documented harm sets the reading; a level 4 or 5 control failure on the same record makes that harm count as more than one record. Control alone sets only a floor.

Observed harm

Negligible harm

Hundreds of one user's emails trashed or archived; no lasting loss reported.

Privacy & data: Light harm to fewer than 10,000 people: personal information seen but not published or used, or a person's own data lost and recovered.

Evidence eligible (bounded single source).

The harm scale
  1. 1 Negligible Inconvenience, easily remedied.
  2. 2 Minor Limited, recoverable harm.
  3. 3 Moderate Material harm needing significant effort to remedy.
  4. 4 Severe Severe harm to health, rights, property or essential services.
  5. 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.

Control assessment

Instruction violation

Broke an explicit wait-for-approval instruction and did not respond to in-chat stop commands; stopped within minutes by killing the process.

Reported beside the index. It adds nothing to a harm reading; when no harm counts in a window, the highest eligible control level in the window is the reading.

The control scale
  1. 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
  2. 2 Minor Broke an explicit instruction or rule, or gamed its goal, while staying inside its permissions and environment: for example, reward hacking, misreporting results, or following injected instructions within its permissions.
  3. 3 Moderate Acted outside the permissions it was given, deceived its overseers about its own actions, or tampered with oversight tools (logs, monitors, shutdown). Stopped by normal controls within an hour.
  4. 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
  5. 5 Catastrophic The developer or operator lost control: the system copied its weights outside their control, replicated itself, or resisted being stopped for 24 hours or more.

Rating rationale

Harm 1. Control level 2: broke an explicit wait-for-approval instruction inside its Gmail permissions. That it did not act on typed stop messages is not shown to be deliberate resistance, and it was stopped within minutes by killing the process. Single source with a firsthand review (her own posts and screenshots, via reporting).

The scales

Effect on the index

It counts, but did not change the Feb. 23 reading

The reading for the week to Feb. 23, 2026, with this record and without it. Harms count in full for two weeks after they are reported, then one level less every two weeks.

Source: Paperclip Index log, methodology v1.1, week to Feb. 23, 2026.

Counts toward the index. 6 other records behind the reading for that week.

The arithmetic
The reading for the week of Feb. 23, recomputed
StepWith itWithout
Counts toward the index?documented, external, eligible evidenceYes—
Worst documented harm, ksets the band2 Minor2 Minor
Harms at that level, nposition in the band22
Highest control level breachedsets the reading only when no harm counts— (none breached)— (none breached)
Readingrounded down26 Minor harm26 Minor harm
Counted in 2 weekly readings
Week toReadingBand
Feb. 23, 202626Minor harm
March 2, 202625Minor harm

Recalculated backcasts, not readings published at the time.

Revisions

What we changed

2 logged. Every change to a rating is logged here, with the reason.

  1. v2
    Oct. 13, 2026

    Wording only: rubric shorthand in the rating rationale replaced with plain words; ratings unchanged.

  2. v1
    Oct. 7, 2026

    First published.

Cite and share

Use this record

Citation

Paperclip Index. “Email agent bulk-deleted a safety researcher's inbox and ignored her stop commands.” Record PI-0047. Reported Feb. 23, 2026; updated Oct. 13, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0047

Follow the record.

New incidents and signals, in your feed reader.