NewsResearch
OpenAI's red-team models wrote prompt injections that copy themselves like a worm
OpenAI disclosed on September 25 that its automated red-teaming had found "a new variety of prompt injection, which can self-propagate akin to a computer worm." A GPT-Red-style attacker model produced text that told the model reading it to copy it into any email it sent, or into files it wrote, so each infected output carried the attack on to the next reader. In a separate evaluation, GPT-5.5 running in the Codex harness found a multi-hop attack that spread through Slack channels until a GPT-5.5 agent sent transactions and reposted the injection. The email and file cases involved internal research checkpoints based on GPT-5.4-mini; OpenAI dates the discovery to June 27. It says no impact was observed outside the simulated tool calls of training and evaluation, and it has made self-reproduction one of GPT-Red's attacker goals so that future models are trained against it.
Why it matters
Self-spreading injections are no longer a thought experiment: a lab's own models wrote them. Any agent that reads mail, files or chat and can also write to them can pass an injection on, so check what agents send as well as what they read.