OpenAI discloses self-replicating prompt injections

Summary

In an alignment report, OpenAI described a new variety of prompt injection that can propagate itself like a computer worm. The finding emerged from its GPT-Red self-play training framework; in one example, an injection arriving by email instructs the agent to copy it into every email it sends. OpenAI says no impact was observed outside simulated tool calls in training and evaluation, and that it is sharing the finding because of its novelty rather than an incident.

Why it matters for our work

As more agents handle email, calendars, and other connectors, self-propagating injections raise the stakes for input validation and least-privilege design in enterprise agent deployments.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.