Papers & Research

UK AI Security Institute and Anthropic publish largest data poisoning study to date

Original title:Examining backdoor data poisoning at scale

AI summary

The UK AI Security Institute, in collaboration with Anthropic and the Alan Turing Institute, published the largest data poisoning study to date. The team tested the same backdoor attack on four models ranging from 600M to 13B parameters and found that just 250 documents were enough to successfully poison the training data of every tested model, with the required number not scaling with model or dataset size. This overturns the prior assumption that proportional poisoning is needed for success and suggests poisoning attacks may be more feasible than previously thought.

Key Takeaways

  • The UK AI Security Institute,in collaboration with Anthropic and the Alan Turing…
  • The team tested the same backdoor attack on four models ranging from 600M to 13B parameters and…
  • This overturns the prior assumption that proportional poisoning is needed for success and suggests…

Quick Facts

Published
Updated
Source
英国 AI Security Institute:Blog(网页)
Category
Papers & Research
AI score
67
Read the original Read the full article at the original source
Why it matters

研究给出与既往假设相反的结论,即毒化所需文档数不随模型规模增长,对理解后门攻击可行性有直接参考价值。

Related entities

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: UK AI Security Institute and Anthropic publish largest data poisoning study to date. October 9, 2025. https://new.nat6.net/en/news/grawjenhg3yqy3qnmbnffiqx7/

View all →