An AI agent that recommends an action and one that executes it carry different operational risks. This briefing examines the chain between the person or organisation granting authority, the agent identity, the tool it invokes and the action the tool can take.
A hypothetical invoice workflow separates reading an invoice, matching a delivery record, proposing a payment and releasing it. Each capability can have a different permission. The paper also considers untrusted instructions in documents and why an application must enforce boundaries around tools and data.
The briefing synthesises public NIST material. It is not a security assessment of a named provider or a guarantee that an agent is safe. The PDF includes the full analysis, method and source links.
Methodology: Desk-based synthesis of public NIST work on AI agent security and identity. The importer example is hypothetical. No provider, model or deployment was independently assessed.
Download report (PDF)