CoT Forgery: The Chain-of-Thought vulnerability in LLM security

This report explains Chain-of-Thought (CoT) Forgery, a prompt-injection attack that inserts fake internal reasoning into LLM prompts to bypass safety controls. Using a banking assistant scenario, it demonstrates how the forgery can trick the model into giving actionable guidance for evading anti-money-laundering controls, and recommends defenses including context sanitization, targeted adversarial testing, and continuous red-teaming.