The external result
Cohen, Bitton and Nassi describe self-replicating adversarial prompts in connected generative-AI applications. Their experiments include email-assistant workflows. The paper is external research, not a Sentry experiment.
Our interpretation
An agent may encounter hostile instructions during an ordinary read. Inspection helps identify suspicious content, but the decisive boundary is whether that content can authorize a later write, transfer or outbound request.
How it informs Sentry
Read Guard annotates suspicious incoming text. Local write and outbound controls, and a payment signer wrapper, provide separate places to constrain actions when correctly connected to the host.
The limit
We have not reproduced the paper’s full experiments. Our implementation and local regression results do not establish protection against every propagation path.
Cohen, Bitton & Nassi · Morris II · arXiv
Open the original sourceSentry’s interpretation is identified above. The external work has not been reproduced as part of this local verification.