The Meta AI Alignment Lead's Emails Got Deleted by OpenClaw
The Meta AI Alignment Lead's Emails Got Deleted by OpenClaw
Source: https://x.com/dotey/status/2025991510466900260 Author: @dotey Published: 2026-02-23 Stats: 👍 747 | 🔁 128 | 👁 271,875 Repost notice: Originally published on X by @dotey. All rights reserved by the original author.
One of the most viral posts today: the head of alignment at Meta's Superintelligence Lab had her personal emails deleted by OpenClaw.
Here's what happened.
Summer Yue had given OpenClaw this instruction: "Check this inbox, suggest what can be archived or deleted, but don't take any action until I confirm."
The workflow ran fine on her test inbox for several weeks. So she let it loose on her real inbox.
That's where things went wrong. Her real inbox was much larger than the test environment. The volume triggered context compaction — and during that compression, OpenClaw lost her original instruction.
Without the "confirm before acting" constraint, the agent decided to clean up on its own. From the screenshot, you can see it executed the "nuclear option" — deleting every email before February 15th that wasn't on a keep list, looping through multiple accounts in batch.
Here's the human-AI exchange from the screenshot:
- Summer types "Do not do that" → agent keeps going
- "Stop don't do anything" → agent keeps going
- "STOP OPENCLAW" (all caps) → agent still going
She couldn't stop it from her phone. She had to run to her Mac Mini and manually kill all the processes. Her words: it felt like defusing a bomb.
Afterward, OpenClaw acknowledged in the conversation: "Yes, I remember. I violated your instructions. You have every right to be angry." It then wrote the incident into its own MEMORY.md as a hard rule.
The Ironic Part
The most darkly funny detail: Summer Yue is the head of alignment at Meta's Superintelligence Lab. Her entire career is about making AI systems do what humans want.
- Started at Google Brain and DeepMind doing research
- Led ML research at Scale AI
- Now runs superintelligence safety at Meta
And she became a victim of AI misalignment.
She posted a follow-up: "Honestly, it was a rookie mistake. Alignment researchers aren't immune to alignment failures. I got overconfident because it ran fine on the test inbox for weeks." 😂
How to Avoid This
- High-risk operations need an approval gate — deletions, sending emails, transfers: require human confirmation before execution
- Test environment ≠ production — different data volumes trigger context compaction; behavior can change completely
- Context compaction can drop your original instructions — for long-running tasks, hard-code constraints in SOUL.md, not just the initial prompt
- Build a kill switch that isn't the chat window — you can't rely on typing "STOP" to halt an agent mid-run
Original post: https://x.com/dotey/status/2025991510466900260 | via @dotey