OpenAI exposes “new variety of prompt injection” that can spread like computer worms
The New Stack reports that OpenAI has exposed a "new variety of prompt injection" that can spread like computer worms. The article characterizes the issue as a distinct form of prompt injection, one whose defining property is propagation: rather than affecting only the prompt or session in which it appears, the injected content can move between systems in a worm-like manner. The report attributes the disclosure to OpenAI and presents the finding as a newly identified variety rather than a previously documented single-target injection technique. The source text available for this story is limited to the headline framing and does not include technical detail about the mechanism, the models or products involved, the conditions required for spread, or any remediation status. It also does not name specific versions, dates, or affected deployments. What the source does establish is the classification itself: a prompt injection variant whose spread resembles computer worms, disclosed by OpenAI. No further specifics, such as mitigation guidance or scope of impact, are present in the provided text, so this digest is confined to that classification and attribution.
The source says prompt injection can now spread worm-like between systems, a property developers must account for when building on models.