Back to today’s briefing

Story perspectives

Study Reveals 250 Malicious Docs Can Corrupt AI Models

10/18/2025

50 7

Summary

Story summary

A joint study by the UK AI Security Institute, the Alan Turing Institute, and Anthropic finds that 250 malicious documents can corrupt AI models and create backdoor vulnerabilities.
The finding challenges the belief that larger models require more poisoned data.
As datasets grow, poisoning risk increases and attackers may exploit a small, fixed number of documents, prompting calls for stronger defenses.

1 of 1

Story summary

A joint study by the UK AI Security Institute, the Alan Turing Institute, and Anthropic finds that 250 malicious documents can corrupt AI models and create backdoor vulnerabilities.
The finding challenges the belief that larger models require more poisoned data.
As datasets grow, poisoning risk increases and attackers may exploit a small, fixed number of documents, prompting calls for stronger defenses.