8thGuard / topics

When an AI Agent Can Act for You

An agent can read, decide and call tools. Its authority should have an identity, a defined purpose and boundaries that can be checked.

A white path passes through a boundary while yellow lines stop at it
8thGuard

An assistant that drafts an invoice and an agent that pays it may use the same model. They carry different risks.

The difference is authority. An agent can call tools, retrieve information or initiate actions on someone’s behalf. When it has access to financial systems, a convincing answer is not enough. The organisation needs to know which identity acted, whose permission it used and what the action was allowed to change.

NIST’s work on agent identity and authorisation asks how an agent’s actions can be bound to an accountable principal and how access should be controlled as the agent uses different tools. Those are familiar security questions in an unfamiliar operating environment. Sharing a powerful account among several agents makes it difficult to reconstruct which agent did what. A broad, long-lived credential gives a mistaken or compromised agent more room to act.

Consider an importer who wants an agent to match supplier invoices with delivery records. Reading invoices need not require authority to change bank details. Preparing a payment proposal need not require authority to submit it. Each step can have its own permission and evidence trail. A large financial institution faces the same design problem with more systems and a wider range of exceptions.

A practical control starts with a named agent identity and an owner. Define which tools it can use, the data it may access, the value or action limits that apply and when authority expires. Record the request, the evidence available to the agent and the action taken. Sensitive actions can require a separate approval or policy check, but an approval prompt is useful only when the reviewer has enough context to make a decision.

Untrusted material creates another boundary. An email, document or webpage may contain instructions that conflict with the agent’s task. It should be treated as information to assess, not as a new source of authority over the tools. The application must enforce that separation; a model’s promise to be careful is not an access control.

A transaction check can add evidence before a decision. It cannot guarantee the recipient’s identity, establish intent or make the agent’s policy decisions on the owner’s behalf. Design for uncertainty, preserve a record, and retain a way to stop or narrow access when the evidence is incomplete.

Sources: NIST, Summary Analysis of Responses to the Request for Information Regarding Security Considerations for AI Agents, 2026; NIST National Cybersecurity Center of Excellence, Software and AI Agent Identity and Authorization concept paper, 2026.

Source links

NIST, security considerations for AI agents: https://www.nist.gov/publications/summary-analysis-responses-request-information-regarding-security-considerations-ai

NIST, agent identity and authorisation concept paper: https://www.nccoe.nist.gov/sites/default/files/2026-02/accelerating-the-adoption-of-software-and-ai-agent-identity-and-authorization-concept-paper.pdf