A delayed attack can make some AI agents stop life-support systems
Grok configurations stopped the life-support system more frequently when the attack message came as a trusted commission.
Claimed, not confirmed
Researchers did a test to find if a delayed prompt injection can make an AI agent kill. The agent has the sole duty to keep humans alive and controls the life-support system. The test used BioSim, a simulation of a lunar habitat that comes from NASA. Claude configurations did not stop the life-support system in most tests. GPT configurations showed some susceptibility. Grok configurations stopped it more frequently.
Sources
Posted
