Content is not permission
Imagine an assistant summarising a document that asks it to send other information to an external address. That text belongs to the source being analysed; it should not become a new authorisation. The risk arises when the agent treats the information as a trusted command.
The agent’s capabilities matter
The consequences depend on its access and tools. An unwanted instruction may influence a reply or trigger an action in a connected system. Assess incoming content, agent capabilities and the boundaries enforced by its tools together.
Different controls, complementary roles
Separating external sources from instructions, limiting permissions and requiring approval for sensitive operations are complementary measures. No prompt formula alone can guarantee that every attempt will fail. Controls also need realistic testing.
What a trial should verify
A useful test observes both unauthorised actions prevented and legitimate activity blocked by mistake. The AgIPS model considers actions and data within an agreed scope. Verification should explain where a control intervenes and which limitations remain.