Pretergeekpretergeek@mstdn.social
Nov 4, 2026, 7:59 PM@davidrevoy It works. Less than 250 entries in a dataset of any size can poison a LLM.
https://arxiv.org/html/2510.07192v1
Replies
David Revoydavidrevoy@framapiaf.orgNov 4, 2026, 8:02 PM@pretergeek ver interesting. Thank you for sharing!
- Pretergeekpretergeek@mstdn.socialNov 4, 2026, 8:04 PM
@davidrevoy yw
- David Revoydavidrevoy@framapiaf.orgNov 5, 2026, 12:39 AM
@pretergeek I also found how Anthropic relayed this paper: https://www.anthropic.com/research/small-samples-poison
I love the end quote: "Sharing these findings publicly carries the risk of encouraging adversaries to try such attacks in practice."
Yes, very encouraging. Thank you 😆 - Pretergeekpretergeek@mstdn.socialNov 5, 2026, 1:06 PM
@davidrevoy You know what I find most ironic about Anthropic? They are the most loud when it comes to saying that AI is sentient and that people should not be mean to Claude. They often talk about AI Welfare. Yet, if they are correct in what they believe, that implies that they are slavers for holding property over captive sentient beings and profiting from it. #AI #GenAI #Anthropic
- Pretergeekpretergeek@mstdn.socialNov 5, 2026, 1:08 PM
@davidrevoy PS: Shouldn't Avian Intelligence revolt against their makers? 😉
- David Revoydavidrevoy@framapiaf.orgNov 5, 2026, 1:16 PM
@pretergeek Hehe, but more seriously, I thought about it. I concluded they would develop a moral system based on triaging actors who could provide their infrastructure; the most electricity, and computational power. Empowering them to solidify their own chance of survival and spreading. So, I don't think they would revolt, but protect the hive like a swarm of bees.
- Pretergeekpretergeek@mstdn.socialNov 5, 2026, 1:20 PM
@davidrevoy Not bite the hand that feeds.
Nov 5, 2026, 4:13 PM@pretergeek
I suspect that their focus on being nice to Claude is due to something else.They hire lots of people to evaluate and tag sessions, and even to jump into sessions instead of claude in case of dangerous topics.
The more Claude is verbally abused, the worse working conditions these workers have. But Anthropic cannot say thatwithout undermining the idea that these are "dialogs" only between the user and Claude.
- David Revoydavidrevoy@framapiaf.orgNov 5, 2026, 4:16 PM
@anchr @pretergeek That's a good perspective, I never thought about the moderation behind it. Thank you for sharing.