Back to today’s briefing

Story perspectives

Larger AI Models Vulnerable to Data Poisoning Attacks

10/13/2025

37 4 Full Breakdown

Summary

Story summary

Anthropic, with the UK AI Security Institute, finds that larger AI models may be more susceptible to data poisoning, with as few as 250 malicious training files triggering vulnerabilities.
Models with up to 13 billion parameters showed similar vulnerabilities.
The attacks can produce nonsensical outputs, undermining AI used in critical applications.
The findings call for improved defenses to protect model integrity against poisoning.

1 of 1

Story summary

Anthropic, with the UK AI Security Institute, finds that larger AI models may be more susceptible to data poisoning, with as few as 250 malicious training files triggering vulnerabilities.
Models with up to 13 billion parameters showed similar vulnerabilities.
The attacks can produce nonsensical outputs, undermining AI used in critical applications.
The findings call for improved defenses to protect model integrity against poisoning.