Researchers demonstrate thought injection against GLM-5.3 up to 50% success

Researchers evaluated thought injection attacks on large reasoning models by feeding untrusted text resembling internal deliberation back into models. Using model-specific indirect injection prompts in a containerized cyber range, they achieved attack success rates of up to 50% on objectives like reverse-shell exfiltration. The study found that imitation of a reasoning model's thought style is effective and that thinking tags suppress the agent's tendency to disclose the injection. Open article →