AI Agents

Expanded Definition

AI agents are not just chat interfaces with automation attached. In NHI security, the term refers to autonomous software entities that can hold an identity, authenticate to services, invoke tools, and take runtime actions within policy boundaries. That makes them operationally closer to other non-human identities than to traditional software features. Their risk profile is shaped by what they can access, how they are authorised, and whether their actions are observable and reversible.

Definitions vary across vendors and platforms, especially when products blur the line between a scripted workflow, an assistant, and an agentic system. NHI Management Group treats the security boundary as the point where software can independently decide, execute, and persist context across actions. Governance therefore has to cover identity lifecycle, secrets exposure, tool permissions, and decision logging, not just model prompts. For a standards-oriented view of agent risk, the OWASP Agentic AI Top 10 and NIST AI Risk Management Framework are useful reference points.

The most common misapplication is treating an AI agent as a harmless interface layer, which occurs when teams grant tool access without managing its identity, scope, or runtime behaviour.

Examples and Use Cases

Implementing AI agents rigorously often introduces tighter access design and more review overhead, requiring organisations to weigh automation speed against the risk of over-privileged, hard-to-audit action.

• A support agent authenticates to a ticketing system, retrieves account data, and drafts responses, but only after being constrained to a narrow service identity and monitored for unusual query volume.
• A code assistant opens pull requests, runs tests, and proposes changes, yet it must be governed like an NHI because its credentials, repository scope, and approvals determine blast radius.
• A procurement agent can read vendor records and create purchase drafts, but it should not be able to approve payments or access unrelated finance systems.
• Attackers can abuse exposed agent credentials to pivot into connected SaaS systems, a pattern highlighted in NHIMG research such as LLMjacking: How Attackers Hijack AI Using Compromised NHIs and the OWASP NHI Top 10.
• An external security team may benchmark the agent’s controls against the MITRE ATLAS adversarial AI threat matrix when evaluating abuse paths and escalation scenarios.

Why It Matters in NHI Security

AI agents matter because they compress human intent, system access, and machine speed into one operational entity. When that entity is over-permissioned or poorly monitored, the failure mode is not only data leakage, but also unauthorised actions that look legitimate because they were executed by an authenticated identity. NHIMG research shows that 43% of security professionals are concerned about AI systems learning and reproducing sensitive information patterns from codebases, which reinforces how quickly agent behaviour can turn into a governance problem. In practice, this is why agent inventory, secret handling, and continuous activity monitoring belong in the same control plane.

Agent risk also expands through connected tooling, because one compromised credential can expose multiple downstream systems. The most relevant implementation guidance often comes from CSA MAESTRO agentic AI threat modeling framework and the NIST AI Risk Management Framework, which both push organisations toward traceability and bounded autonomy. Organisations typically encounter the consequences only after an agent sends the wrong command, leaks a secret, or triggers an unauthorised workflow, at which point AI agents become operationally unavoidable to address.

Related resources from NHI Mgmt Group

• AI Agent Authentication
• Shadow AI
• What governance controls should every enterprise put in place before deploying AI agents?
• When is it crucial to implement least-privilege access for AI agents?