Insight · Security
Agentic AI Security & Controls
Agentic systems change enterprise risk because software can select tools, access data, retain context, and take actions with less direct human intervention. Security must govern what an agent is allowed to know, decide, access, and do.
Executive context
Why this matters
- An agent may combine model reasoning with identity, memory, tools, data access, delegated authority, and connections to enterprise systems.
- Traditional application controls remain necessary, but they must be extended to govern instructions, tool selection, action boundaries, evaluation, and human intervention.
- Leaders need a control model that connects AI-specific risk to existing security, risk, architecture, and operational-resilience responsibilities.
Core content
The operating considerations behind the idea
Identity before autonomy
Every production agent needs an explicit identity and an accountable owner. Access should be granted to that identity—not inherited broadly from a developer, service account, or end user.
- Unique workload or agent identity
- Least-privilege authorization
- Short-lived credentials where practical
- Separation between development, testing, and production
Bound knowledge and tools
Agent security depends on controlling both the context an agent can retrieve and the tools it can invoke. Data boundaries, retrieval authorization, tool allowlists, and policy enforcement should be designed together.
- Identity-aware data retrieval
- Approved tool inventory
- Explicit input and output boundaries
- Third-party model and service dependencies
Govern consequential actions
Not every action should be autonomous. Organizations should define which decisions require deterministic checks, human approval, escalation, or prohibition.
- Risk-based approval thresholds
- Human authorization for consequential actions
- Deterministic validation at critical boundaries
- Containment and kill mechanisms
Observe and evaluate the complete chain
Monitoring must cover more than model output. Useful evidence includes instructions, retrieved context, tool calls, policy decisions, approvals, actions, outcomes, and exceptions.
- Traceable activity logs
- Quality and security evaluations
- Anomaly and misuse detection
- Incident response and recovery paths
Connect to enterprise controls
Agentic security should extend existing enterprise security practices rather than become an isolated control program. The Studio maps this work to Govern, Identify, Protect, Detect, Respond, and Recover outcomes in the NIST Cybersecurity Framework where appropriate.
- Governance and risk ownership
- Architecture and threat modeling
- Identity, data, and platform controls
- Assurance, resilience, and continuous improvement
Leadership questions
Questions worth asking before the next decision
- 01Which agents can take actions that affect members, employees, money, systems, or regulated decisions?
- 02Who owns each agent and accepts residual risk?
- 03What information can the agent retrieve, retain, or disclose?
- 04Which tools and systems can it invoke, and under whose authority?
- 05Where are human approval and deterministic checks required?
- 06Can the organization reconstruct, interrupt, and contain an agent's activity?
Practical implications
What leadership and delivery teams should consider
- Inventory agents as governed enterprise identities and assets.
- Classify use cases by action authority, data sensitivity, reversibility, and business consequence.
- Design authorization, observability, evaluation, and containment before production release.
- Map AI-specific controls into existing security, vendor-risk, incident, and assurance processes.
- Use evidence from controlled implementations to refine policy and risk thresholds.
Related Studio assets
Continue into the supporting frameworks and applied work
Enterprise Security Framework
See how agentic controls connect to the broader enterprise security operating model.
Explore the Security FrameworkEnterprise Agentic Transformation
Understand how working agents fit within governance, platforms, people, and measurable outcomes.
Explore the Agentic FrameworkResponsible AI Governance
Connect security controls to decision rights, lifecycle oversight, and human accountability.
Explore Responsible AI GovernanceRelated advisory service
AI Governance & Responsible Adoption
Clarify accountability, risk tiers, policies, controls, and oversight for AI and agentic systems.