Can AI Agents Hide Backdoors? Anthropic Research Reveals a New Data Poisoning Threat
29 October 2026
New Anthropic-linked research shows how poisoning just 0.1% of AI training data can create hidden backdoors in coding agents that survive clean training.