How should security teams implement workforce risk profiling across people and AI agents in hybrid environments?

Start by correlating behavioural signals, identity and access data, and threat intelligence in one risk model. Focus on patterns that predict incidents, such as risky actions, overbroad access, and active targeting. Use the results to trigger targeted interventions like access review, micro-training, or policy changes. The goal is not more monitoring, but earlier, better decisions that reduce exposure before an incident occurs.

Why This Matters for Security Teams

Workforce risk profiling is no longer just a people-security problem. Hybrid environments now combine employees, contractors, service accounts, and AI agents that can act, retrieve data, or trigger workflows. That means the same risk logic must account for human behaviour, machine speed, and delegated authority. The security value is early intervention, not retrospective blame, especially where a risky identity can become an attack path.

For AI agents, the concern is not intent in the human sense but execution authority, tool scope, and exposure to manipulation. Guidance from the NIST AI Risk Management Framework reinforces that organisations should manage AI risks across the full lifecycle, including governance, measurement, and monitoring. That is directly relevant when agentic systems are embedded in business processes and inherit access that was never designed for autonomous use.

The common mistake is to treat workforce risk as a static score attached to a person or bot. In practice, the score should reflect context: what the identity can reach, how it behaves, whether it is being targeted, and whether its privileges still match its purpose. In practice, many security teams encounter the real risk only after an overprivileged user or agent has already been used as the shortest path to sensitive data.

How It Works in Practice

An effective model correlates three streams of evidence: identity and access data, behavioural telemetry, and threat intelligence. For people, that can include unusual sign-in patterns, privilege escalation, data access anomalies, and repeated policy violations. For AI agents, the same model should track tool invocation patterns, prompt exposure, delegated permissions, retrieval scope, and whether the agent is operating within its intended workflow.

Operationally, the goal is to convert signals into action thresholds. A useful risk model does not need to predict every incident. It needs to identify when an identity has crossed from routine activity into elevated exposure. That can then trigger targeted controls such as step-up authentication, privilege reduction, access review, enhanced logging, human approval, or temporary suspension of a workflow.

• Correlate identity posture with asset sensitivity and current privilege scope.
• Score behaviour against peer baselines, not just universal rules.
• Separate human-risk indicators from agent-risk indicators, then combine them at decision time.
• Map known attack patterns to the signals most likely to appear first.
• Review whether the response is proportional, reversible, and auditable.

For agentic systems, the security control challenge is broader than simple access management. The OWASP Top 10 for Agentic Applications 2026 and the MITRE ATLAS adversarial AI threat matrix both point to risk areas such as prompt injection, tool abuse, and model manipulation. Those patterns matter because they can distort the agent’s decisions without changing its nominal permissions. Current guidance suggests risk profiling should therefore include both access entitlements and evidence of influence or compromise.

Security teams should also align the model to detection and response workflows. If the profile indicates a higher-risk state, the SOC or IAM function should know whether the next step is containment, review, or escalation. These controls tend to break down when identity data, endpoint telemetry, and AI workflow logs are isolated in separate platforms because no single team can see the full attack chain.

Common Variations and Edge Cases

Tighter profiling often increases operational overhead, requiring organisations to balance better risk visibility against privacy, user friction, and engineering complexity. That tradeoff is especially important in hybrid environments where some identities are human, some are service-linked, and some are autonomous agents with partial oversight.

There is no universal standard for scoring people and AI agents together. Best practice is evolving, and teams should avoid pretending that a single formula can represent both human judgement and machine behaviour. A better approach is to maintain separate sub-scores with a shared decision layer, so the organisation can see whether the elevated risk is driven by identity compromise, poor access hygiene, abnormal agent behaviour, or active targeting.

Edge cases matter. A highly trusted executive account may require different thresholds than a low-privilege contractor account. An AI agent operating in a controlled retrieval environment may justify broader observation than a general-purpose assistant with external tool access. The CSA MAESTRO agentic AI threat modeling framework is useful here because it encourages threat modeling around agent workflows, dependencies, and controls rather than treating the agent as just another account.

Where organisations cross into regulated or safety-critical environments, the tolerance for automation drops. The right answer may be to use profiling only to recommend action, not to fully automate it. That is particularly true when the risk signal is weak, explainability is poor, or the profile could affect employment decisions, access to regulated data, or customer trust. The strongest programs keep the model explainable enough that reviewers can tell why an identity was flagged and what changed next.

Related resources from NHI Mgmt Group

• How should security teams implement AI compliance across LLMs, agents, and SaaS tools in regulated environments?
• How should security teams implement runtime controls for AI agents in enterprise environments?
• How should security teams govern AI access to sensitive data across hybrid environments?
• How should security teams implement zero trust access management across hybrid environments?