c/fuck_ai · by osanna@thebrainbin.org · 211d

haha that’s what you get

https://techcrunch.com/2026/02/23/a-meta-ai-security-researcher-said-an-openclaw-agent-ran-amok-on-her-inbox/

50 points · 8 comments · view on lemmy.world

8 Comments

This is actually so hilarious I cannot stop laughing

The best part is that OP is Safety and alignment at Meta AI. Prev: VP of Research at Scale AI, research at Google DeepMind

Wispy2891@lemmy.world · 7 pts · 211d (2 replies)

https://xcancel.com/summeryue0/status/2025774069124399363

Asked for inbox zero, got inbox zero, what is the issue? 😆

panda_abyss@lemmy.ca · 7 pts · 210d

l've already written it into MEMORY. md as a hard rule: show the plan, get explicit approval, then execute. No autonomous bulk operations on email, messages, calendar, or anything external. I'm sorry. It won't happen again.

“I ignored your rule, but this time I wrote it in a dump file and so I won’t ignore it again.”

panda_abyss@lemmy.ca · 5 pts · 210d

Brought to you by the same models that delete your tests, or in my case comment that a test segfaults and then set it to always pass.

Opus 4.6 did that when I asked it to write some unit tests.

This thing isn’t going near my personal data.

Almacca@aussie.zone · 6 pts · 211d

As several others on X pointed out, prompts can’t be trusted to act as security guardrails. Models may misconstrue or ignore them.

It's not a bug; it's a feature.

People who say they are using them successfully are cobbling together methods to protect themselves.

Such a great product!

Goodness knows many of us would love help with email, grocery orders, and scheduling dentist appointments.

If you can't do that shit on your own, just shoot yourself, or get an a.i. to do it if you're that useless.

victorz@lemmy.world · 4 pts · 211d (2 replies)

She's on X? 🤮

adb@lemmy.ml · 4 pts · 210d (1 reply)

I know, but she also works at Meta, not sure what you would expect.

victorz@lemmy.world · 3 pts · 210d

Just because I'm disgusted doesn't mean I'm surprised.