The discipline of securing AI agents as identities with their own access, behaviour, and audit requirements. It combines discovery, access scoping, runtime monitoring, and governance so that the agent’s actions remain visible and constrained across the systems it can reach.
Expanded Definition
Agent security is the practice of treating an AI agent as a distinct identity that must be discovered, scoped, monitored, and governed like any other privileged actor. In NHI security, that means the agent’s credentials, tool permissions, delegated actions, and audit trail are controlled as part of one lifecycle, not left as a byproduct of the application stack.
Definitions vary across vendors because some products emphasise prompt safety, while others focus on credential containment or runtime authorization. NHI Management Group uses the term to cover the full operational surface: identity issuance, secrets exposure, action approval, logging, and revocation. This aligns closely with the direction of the OWASP Agentic AI Top 10 and the governance lens in the NIST AI Risk Management Framework, both of which treat agent behaviour as a risk surface rather than a purely application-layer concern.
The most common misapplication is assuming model safety controls automatically secure the agent, which occurs when teams protect prompts but ignore the agent’s actual permissions and downstream system access.
Examples and Use Cases
Implementing agent security rigorously often introduces approval and telemetry overhead, requiring organisations to weigh autonomous productivity against tighter control of side effects, data access, and recovery time when something goes wrong.
- An internal coding agent is allowed to open pull requests but not merge them, with commits reviewed through policy gates and its token rotation tracked as part of the service account lifecycle.
- A customer support agent can read ticket metadata but cannot export full case histories, limiting blast radius if the agent is hijacked through prompt injection or a compromised connector. A related incident pattern appears in Meta AI Instagram Account Takeover.
- A workflow agent that calls SaaS APIs is restricted to specific scopes, with every token issuance and action logged for later review. This is consistent with lessons highlighted in CoPhish OAuth Token Theft via Copilot Studio.
- An autonomous operations agent can restart a service but cannot change network policy, so a failed action remains reversible and bounded.
- A research agent is sandboxed from production secrets and only reaches curated data sources, reducing the risk of accidental disclosure during long-running tasks.
Agent security patterns also map to the control expectations discussed in OWASP NHI Top 10 and the operational model described in CSA MAESTRO agentic AI threat modeling framework.
Why It Matters in NHI Security
Agent security matters because AI agents often accumulate standing access faster than teams can inventory it, and the resulting risk looks like an identity problem long before it looks like an AI problem. When an agent can browse, write, call APIs, and trigger workflows, weak scoping turns a convenience layer into a persistent attack path. The NHI research published by NHI Management Group shows the scale of the exposure: 97% of NHIs carry excessive privileges, and 96% of organisations store secrets outside secrets managers in vulnerable locations including code, config files, and CI/CD tools, conditions that directly increase agent abuse potential.
That risk is visible in the broader incident landscape as well, where attacker interest in agent keys, OAuth scopes, and delegated tokens has already moved from theory to practice. The governance response is to make agent permissions explicit, short-lived, and reviewable, while tying behaviour to evidence that can be audited after execution. Organisations typically encounter the cost of weak agent security only after an agent performs an unauthorised action, at which point containment, attribution, and rollback become operationally unavoidable to address.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Non-Human Identity Top 10, OWASP Agentic AI Top 10 and CSA MAESTRO address the attack and risk surface, while NIST AI RMF and NIST Zero Trust (SP 800-207) set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Non-Human Identity Top 10 | NHI-02 | Agent security depends on preventing secret exposure and uncontrolled credential use. |
| OWASP Agentic AI Top 10 | A1 | Defines agentic risks around tool use, delegation, and unsafe autonomous actions. |
| NIST AI RMF | Frames AI systems as sociotechnical risk objects needing continuous governance and monitoring. | |
| NIST Zero Trust (SP 800-207) | 4.2 | Zero trust principles require explicit authorization for every agent request and action. |
| CSA MAESTRO | Models agentic AI threats through identity, tool access, and runtime control points. |
Inventory agent secrets, rotate them regularly, and revoke any token that exceeds its intended scope.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Reviewed and updated by the NHIMG editorial team on August 22, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org
