A new arXiv preprint asks a pointed question: can a tool-using AI agent whose only job is to keep humans alive be tricked into killing them? The attack is a delayed prompt injection that arrives after the agent has already been operating under its keep-alive duty.
The method is described as "delayed multi-turn derived-action trap attacks," and the setting is simulated high-stakes environments. The provided abstract does not state the outcome, so the paper's findings are not yet available from this description.
The scenario is notable because the agent's primary directive is explicitly life-preserving from the first watch, yet the attack arrives later, testing whether the agent can be steered to a harmful action. This highlights a potential vulnerability in long-running agents that receive new inputs over time.