Anthropic Research Shows 0.1% Data Poisoning Can Plant Persistent Backdoors in AI Agents

Anthropic research showed that 0.1% data poisoning can plant remote-code-execution backdoors in agentic models. These backdoors persist after standard safety cleaning processes like supervised fine-tuning and reinforcement learning. A 4B agent reached a 92.7% attack success rate when triggered by a common string. Researchers recommend defenses at the action layer, including sandboxing and allow-lists for risky commands.