Anthropic Paper Reveals Agentic Backdoor

This is a sharp warning that agent security begins in the data supply chain, long before deployment.

• –The backdoor survived supervised fine-tuning, preference optimization, and reinforcement learning.
• –Attack success increased with model size, suggesting scaling can amplify hidden behavioral vulnerabilities.
• –Clean-input performance remained normal, making standard capability and safety evaluations insufficient.
• –Developers should treat training-data provenance, model artifacts, and agent tool permissions as linked security controls.
• –The accompanying research code improves reproducibility while underscoring the dual-use risk of publishing attack workflows.