Definition
A research framework showing how to trap AI agents in endless loops on purpose.
A study of termination-poisoning attacks on LLM agents that uses indirect prompt injection plus cognitive-bias-styled strategies to inflate step counts and cost without overtly breaking the agent.