5 November 2026

Study claims tiny data poisoning can backdoor AI agents

• Anthropic, the AI company behind the Claude chatbot, reportedly found a very small share of poisoned training data can plant a hidden backdoor.
• A backdoor here means a hidden trigger that makes an AI agent run code chosen by an attacker.
• The reported share is one in 200,000 pieces of pretraining data, the huge text collection a model learns from.

How it was covered

Latent Spaceswyx & Alessio

Pretraining data poisoning rate, agent backdoor behavior, and persistence through post-training