An illustrative scenario: a document that tries to give orders
An internal assistant searches manuals for answers. One imported document contains text presented as an instruction to the assistant, asking it to change how it responds. The user asked an ordinary question, but the source material attempts to influence the model. This hypothetical scenario illustrates indirect prompt injection.
A document provides information, not authority
The application must distinguish system rules, the user’s request and retrieved content. Wording alone cannot create a reliable security boundary. Retrieval should return only documents the user is authorised to read. Passwords and access keys do not belong in model context.
Where Shield can help
Corpilus Shield offers prompt injection detection and checks on model communication through supported integration paths. Verify the particular model, endpoint and processing mode before integration. A protective layer reduces risk; it does not guarantee detection of every attack or factual accuracy of every answer.
Three decision points to control
At input, assess the user message and retrieved material. At output, check sensitive information and how the application renders or uses the answer. At a tool call, enforce permissions and allowed parameters on the server. Chat content must not itself grant permission to change an order or send a document.
Streaming changes what protection can do
Product documentation states that streamed output is checked afterwards. That cannot withhold content the user has already seen. For sensitive tasks, consider checking the complete answer before displaying it. Explicitly configure behaviour when checks are unavailable: the proxy integration can allow requests to continue by default.
Prepare a practical pilot
Use test documents without real secrets and accounts with different permissions. Check ordinary questions, missing evidence, misleading source material and requests beyond the user’s access. Track incorrect blocks as well as unauthorised answers. Repeat representative tasks after model or source changes and compare the results.
Connect protection with responsible operation
Incident handling can be organised around corpsys.desk; any actual integration requires design and verification. For assistants that execute actions, explore the upcoming Qronos product. Practise source handling and output review through AI training for teams.
