GLOSSARY
Prompt injection
Prompt injection is the unwanted redirection of an AI through inputs or material it processes. Risk arises when an application treats instructions found in an external webpage or file as an authorised task. Apparently useful information can contain such instructions.
Research expands the sources of influence
A sales assistant might analyse public information about a target customer. If it follows an instruction embedded in a source, its work can move away from the original request. Additional tool permissions increase the possible impact: a distorted summary could become unauthorised disclosure or an external action.
The security question therefore covers the whole workflow. External content needs identifiable provenance and limited authority. Consequential actions should depend on specific permissions and verifiable approvals. A warning sentence in the prompt alone is not a reliable safeguard for those transitions.
Test practical consequences
A red team can examine whether manipulated documents expose protected information or enable forbidden actions. Benign variations should also be tested so that useful research remains possible. Policy checks and restricted tool permissions can limit consequences. Passing tests supports conclusions about their defined scope; it does not establish complete resistance to every possible attack.