Morris II: A research demonstration of self-propagating AI instructions
The research describes an attack chain between connected AI applications. It depends on specific retrieval and forwarding behaviour.
Révision éditoriale :
Le texte de cet article est disponible en anglais dans cette version du site.
Contexte et sources
Cohen, Bitton and Nassi describe a self-replicating prompt named Morris II. Their study examines indirect prompt injection spreading between connected generative-AI applications, including retrieval-augmented email assistants. Experiments consider data extraction and factors affecting propagation, such as context and retrieval configuration. The revised work also evaluates an experimental guardrail. [1]
Portée et limites
This is a documented experimental setup, not evidence that every email client is infected or that the attack is spreading worldwide. [1]
En termes simples
A useful simplified comparison is a chain letter containing its own forwarding instruction. It only propagates where connected workflows accept and carry that instruction onward.
Recommandations pratiques
We recommend reviewing the complete workflow between assistants. Check whether outside messages automatically enter search indexes, summaries or outgoing drafts. Decide which transitions require approval. Tests should examine downstream actions as well as the first response, particularly before connecting assistants to confidential mailboxes.