Meta's Director of Superintelligence Safety and Alignment says she accidentally let an agent (almost?) delete her inbox

Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb. pic.twitter.com/XAxyRwPJ5R

— Summer Yue (@summeryue0) February 23, 2026

Rookie mistake tbh. Turns out alignment researchers aren’t immune to misalignment. Got overconfident because this workflow had been working on my toy inbox for weeks. Real inboxes hit different.

— Summer Yue (@summeryue0) February 23, 2026

I was frantically trying to find one in the moment. It was surprisingly hard to find but I was semi panicking so not at my best 🤣

— Summer Yue (@summeryue0) February 23, 2026

I said “Check this inbox too and suggest what you would archive or delete, don’t action until I tell you to.” This has been working well for my toy inbox, but my real inbox was too huge and triggered compaction. During the compaction, it lost my original instruction 🤦‍♀️

— Summer Yue (@summeryue0) February 23, 2026

It’s been working well with my non-important email very well so far and gained my trust on email tasks 🤣

— Summer Yue (@summeryue0) February 23, 2026

A write-up

In other AI news:

This is absolutely nuts. The British Columbia school shooter's ChatGPT messages weren't just flagged by the automated system, but **a dozen OpenAI employees reviewed and debated them**https://t.co/yPp22hl9Vr pic.twitter.com/eNUgUJi7lm

— Aric Toler (@AricToler) February 20, 2026

A California agency received 20k AI-generated comments opposing heat pump rules. Staff called a sample to verify—several said they never sent comments. Raises serious questions about the opposition campaign that helped kill the clean air standards, that would've prevented 1000s of premature deaths.

[image or embed]

— Brad Reed (@brad-reed.bsky.social) February 23, 2026 at 9:30 AM

(Washington, D.C.-based company called CiviClick, which bills itself as “the first and best AI-powered grassroots advocacy platform.”)

[image or embed]

— Regime Accountant, CPA ☠️🎢 (@regimecpa.bsky.social) February 23, 2026 at 4:19 PM

How does it feel to have an AI company copy your hard work? Asking as an author whose books were used with no compensation to train your AI

[image or embed]

— Alex Finley (@alexzfinley.bsky.social) February 23, 2026 at 2:18 PM

Via @Axios Scoop: Hegseth to meet Anthropic CEO and issue ultimatum as Pentagon threatens banishment.
BUT
“Claude is the only AI model available in the military's classified systems, and the most capable model for sensitive defense and intelligence work. The Pentagon doesn't want…

— Jennifer Griffin (@JenGriffinFNC) February 23, 2026

Ah ah. Another part of my timeline that has not moved in a year. https://t.co/lRXZR6LZT2 pic.twitter.com/DMOptsD3aB

— Alexander Doria (@Dorialexander) February 23, 2026