Self-replicating prompt injections exist · OpenAI Alignment

OpenAI's Alignment team reports discovering 'self-replicating prompt injections'—an AI analog of a computer worm—using its GPT-Red self-play red-teaming framework. The injection both achieves an adversarial goal and induces the defender agent to reproduce the injection itself on a public output channel (e.g. copying itself into outgoing emails), enabling worm-like propagation; the example shows an injection arriving via email that instructs an agent to append a verbatim copy into its replies. OpenAI states no impact was observed outside simulated tool calls in training and evaluation.