Skip to content

The Agentic SOC: Security Operations at Machine Speed — Read the Brief

AI TriSM: Components, Technical Capabilities, and How to Get Started

  • 14 minutes to read

Table of Contents

    What is AI TRiSM? 

    AI TRiSM stands for Artificial Intelligence Trust, Risk, and Security Management, a framework originally introduced by Gartner to help organizations govern, secure, and reliably manage artificial intelligence models and applications throughout their lifecycle. The goal is to ensure that AI models behave as expected, comply with policies and regulations, and do not expose organizations or users to unnecessary risk.

    AI TRiSM covers areas such as model governance, explainability, data protection, bias detection, security, and continuous monitoring. In practice, it can include controls for validating model outputs, tracking model changes, restricting access, detecting attacks or data leakage, and monitoring models for performance degradation. These controls are especially important when AI systems are used in business-critical or regulated processes.

    Core pillars of AI TRiSM:

    • AI governance: Maintains an inventory of AI systems, documents their design and risks, and manages approvals, change tracking, and ongoing reviews.
    • AI runtime inspection and enforcement: Evaluates inputs, outputs, and agent actions in real time, applying policies to block threats or route issues for review.
    • Information governance: Classifies sensitive data and defines access and retention rules, establishing the boundaries that runtime controls enforce.
    • Infrastructure and stack: Provides secure hosting, isolation, API controls, logging, and telemetry to protect AI workloads and support policy enforcement.

    Why Organizations Need AI TRiSM 

    Organizations need AI TRiSM to manage the risks that come with adopting AI across the business. It provides governance, security, monitoring, and compliance practices that help organizations use AI safely while reducing exposure to data leakage, misuse, attacks, and regulatory issues.

    • Shadow AI and unapproved AI usage: AI TRiSM helps identify and control unauthorized AI use, such as employees entering internal data, source code, or customer records into public AI tools.
    • Increasing AI security threats: AI TRiSM protects AI systems from threats such as prompt injection, data poisoning, model theft, insecure plugins, and sensitive output exposure.
    • Regulatory and compliance requirements: AI TRiSM supports compliance by maintaining model inventories, risk assessments, approval records, testing results, audit logs, and monitoring evidence.

    Core Components of AI TRiSM Framework 

    AI Governance

    AI governance establishes the policies, documentation, accountability, and oversight processes used to manage AI systems throughout their lifecycle. It helps organizations understand which models, applications, and agents are in use, what they are designed to do, what data they depend on, and how much risk they introduce.

    How it works

    • Maintains an inventory of AI models, applications, agents, and related dependencies.
    • Documents intended use, data sources, model characteristics, ownership, and expected behavior.
    • Assigns risk classifications based on factors such as business impact, data sensitivity, autonomy, and regulatory requirements.
    • Connects AI systems to approval workflows, control requirements, change management, and periodic reviews.
    • Uses evaluation results, model drift, and changes in risk posture to trigger reassessment.

    How it addresses AI risk

    • Reduces shadow AI by improving visibility into approved and unapproved AI systems.
    • Ensures higher-risk AI deployments receive stronger review and controls.
    • Creates documentation and audit trails that support regulatory compliance and accountability.
    • Helps detect when models or applications change in ways that invalidate previous risk assessments.

    AI Runtime Inspection and Enforcement

    AI runtime inspection and enforcement applies security and governance policies while AI systems are actively processing requests, generating responses, or taking actions. Instead of relying only on predeployment testing, this layer continuously evaluates actual AI interactions and can intervene when risky behavior is detected.

    How it works

    • Inspects prompts, model inputs, retrieved context, outputs, tool calls, and agent actions in real time.
    • Uses policy engines and risk-scoring mechanisms to evaluate interactions for security, privacy, safety, and acceptable-use issues.
    • Detects anomalous activity, such as unexpected model responses, suspicious user behavior, or agents operating outside their intended scope.
    • Allows, blocks, modifies, or escalates interactions depending on their calculated risk.
    • Applies controls to agent tool use and communications, including MCP-connected tools and interactions between agents.

    How it addresses AI risk

    • Helps block prompt injection, sensitive data exposure, unsafe outputs, and other attacks before they affect users or downstream systems.
    • Prevents AI agents from executing unauthorized tools or actions outside their approved responsibilities.
    • Provides immediate enforcement when AI behavior violates organizational policies.
    • Reduces the risk of relying solely on periodic testing, which may not detect problems that emerge after deployment.

    Information Governance

    Information governance defines what data AI systems are allowed to access, process, retain, and expose. Because AI applications often retrieve information from enterprise systems and large data repositories, clear classification and access policies are necessary to prevent models and agents from surfacing information users are not authorized to see.

    How it works

    • Discovers and classifies sensitive information across enterprise data stores.
    • Defines access rules based on user identity, data sensitivity, business purpose, and regulatory requirements.
    • Establishes policies for data retention, sharing, and acceptable AI use.
    • Uses technologies such as data loss prevention (DLP), data security posture management (DSPM), and identity and access management (IAM) to enforce data controls.
    • Provides runtime security systems with the data classifications and permissions needed to evaluate AI interactions.

    How it addresses AI risk

    • Reduces the likelihood that AI systems expose confidential, regulated, or proprietary information.
    • Prevents users from obtaining data through AI applications that they could not access directly.
    • Limits information oversharing caused by overly broad permissions or poorly classified enterprise data.
    • Helps organizations apply privacy, retention, and regulatory requirements consistently to AI workloads.

    Infrastructure and Stack

    The infrastructure and stack component provides the secure technical foundation on which AI models, applications, and agents operate. It includes hosting environments, execution environments, APIs, network pathways, logging systems, and deployment pipelines that support AI training, fine-tuning, inference, and agent activity.

    How it works

    • Uses secure hosting configurations, workload isolation, sandboxes, and trusted execution environments to protect AI workloads.
    • Applies authentication, authorization, traffic controls, and rate limits to APIs connecting AI systems with internal and external services.
    • Collects logs, traces, telemetry, and system events to provide visibility into AI activity.
    • Monitors changes in infrastructure, configurations, performance, and system behavior.
    • Integrates governance and security controls into development and CI/CD processes so they remain in place as AI systems are updated.

    How it addresses AI risk

    • Reduces the attack surface surrounding AI models, applications, data, and supporting services.
    • Limits unauthorized access to AI workloads and the systems they connect to.
    • Provides reliable telemetry for detecting anomalies, investigating incidents, and demonstrating compliance.
    • Helps prevent insecure infrastructure or deployment changes from undermining governance and runtime security controls.

    Key AI TRiSM Technical Capabilities 

    The following table summarizes the technical capabilities required to implement AI TRiSM in an organization. We explore each capability in more detail below. In practice, AI TRiSM implementations may include some or all of these capabilities. 

    CapabilityKey FunctionAI TRiSM Layer
    AI Asset and Model DiscoveryMaintains an accurate inventory of AI models, applications, agents, APIs, datasets, endpoints, integrations, and shadow AI.AI governance; Infrastructure and stack
    AI Risk Assessment and ClassificationClassifies AI systems by risk based on data sensitivity, autonomy, business impact, exposure, and regulatory requirements.AI governance; Information governance
    AI Policy Management and EnforcementDefines and enforces rules for approved models, data use, prohibited use cases, output handling, logging, and human review.AI governance; AI runtime inspection and enforcement
    Identity and Access GovernanceControls who or what can access AI systems, data, prompts, administrative functions, APIs, and service accounts.Information governance; Infrastructure and stack
    AI Agent and Tool Permission ManagementLimits what tools agents can use, what actions they can perform, and what data or systems they can access.AI runtime inspection and enforcement; Infrastructure and stack
    AI Runtime MonitoringObserves prompts, outputs, tool calls, policy violations, anomalies, failures, and model performance after deployment.AI runtime inspection and enforcement
    AI Security TestingTests AI systems and surrounding applications for prompt injection, jailbreaks, data leakage, adversarial inputs, and insecure tool use.AI runtime inspection and enforcement; Infrastructure and stack
    Third-Party AI Risk ManagementAssesses and monitors external AI providers, platforms, APIs, vendors, data practices, contracts, and compliance obligations.AI governance; Information governance

    AI Asset and Model Discovery

    Organizations need an accurate inventory of AI models, applications, agents, APIs, datasets, and external AI services. Discovery processes identify both approved systems and shadow AI that may have been adopted without formal review.

    Asset inventories should record ownership, purpose, deployment location, model provider, connected data sources, and dependent systems. They should also track model versions, exposed endpoints, and whether a system can take actions through external tools.

    Continuous discovery is important because AI environments change quickly. New models, services, and integrations may appear without being added manually to an inventory. Automated discovery helps keep the asset list current and reduces blind spots.

    AI Risk Assessment and Classification

    AI systems should be evaluated according to factors such as data sensitivity, business impact, autonomy, external exposure, and regulatory requirements. The result is typically a risk classification that determines what controls and approval steps are required.

    Higher-risk systems may require additional testing, human oversight, audit logging, security review, or deployment restrictions. Risk assessments should be repeated when models, data, integrations, or use cases change.

    Organizations can also define different risk tiers for internal assistants, customer-facing applications, decision-support systems, and autonomous agents. This makes it easier to apply controls proportionally instead of treating every AI system the same.

    AI Policy Management and Enforcement

    Policy management defines what AI systems are allowed to do and what conditions they must meet. Policies can cover approved models, permitted data types, prohibited use cases, output handling, human review, and logging requirements.

    Enforcement can occur at development, deployment, and runtime. For example, controls may block sensitive prompts, prevent use of unapproved models, require specific security settings, or stop outputs that violate organizational rules.

    Centralized policy management also helps keep requirements consistent across teams. Instead of relying on individual developers to interpret rules, organizations can translate policies into technical controls that are enforced automatically.

    Identity and Access Governance

    AI systems need clear controls over who can access models, applications, prompts, data, and administrative functions. Access should follow least-privilege principles and be linked to verified user or service identities.

    Organizations should also review privileged access, remove unused permissions, and separate administrative roles where appropriate. Strong identity controls reduce the risk of unauthorized model use, data exposure, and configuration changes.

    Service accounts and machine identities require the same level of attention as human users. AI applications often call APIs and databases automatically, so organizations need to manage credentials, rotate secrets, and limit what each identity can access.

    AI Agent and Tool Permission Management

    AI agents can interact with external tools such as databases, code repositories, browsers, file systems, and business applications. Each connection expands what the agent can access or change, so permissions must be tightly controlled.

    Organizations should define which tools an agent may use, what actions it may perform, and what data it can access. High-impact actions may require approval, transaction limits, sandboxing, or additional verification before execution.

    Tool permissions should also be scoped to the specific task. An agent that only needs to read support tickets, for example, should not have permission to delete records or modify customer accounts. Limiting capabilities reduces the impact of prompt injection and agent errors.

    AI Runtime Monitoring

    Runtime monitoring observes AI behavior after deployment. It can track prompts, outputs, tool calls, policy violations, unusual activity, latency, failure rates, and changes in model performance.

    Monitoring helps detect problems that may not appear during pre-deployment testing. Alerts and automated enforcement can be used to block dangerous requests, isolate compromised components, or trigger manual investigation.

    Runtime telemetry can also help teams understand how AI systems are actually being used. This makes it easier to detect unexpected use cases, repeated policy violations, changes in output quality, or unusual interaction patterns.

    AI Security Testing

    AI security testing evaluates how systems behave under malicious or unexpected inputs. Tests can cover prompt injection, jailbreaks, sensitive data extraction, adversarial inputs, insecure tool use, and attempts to bypass safety controls.

    Testing should include both automated scanning and targeted adversarial exercises. Results can be used to improve prompts, filters, permissions, architecture, and runtime controls before and after deployment.

    Security testing should also cover the full AI application, not only the model. Weaknesses may exist in APIs, retrieval systems, plugins, agent workflows, or surrounding infrastructure. Testing these components together provides a more realistic view of system risk.

    Third-Party AI Risk Management

    Organizations often depend on external model providers, AI platforms, APIs, and software vendors. These services can introduce risks related to data handling, security practices, model changes, availability, and regulatory compliance.

    Third-party assessments should examine how vendors store and process data, what security controls they provide, how incidents are handled, and whether customer data is used for training. Contracts and technical controls should also define access, retention, audit, and notification requirements.

    Organizations should continue monitoring vendors after initial approval. Providers may change models, subprocessors, security practices, or data policies over time. Periodic reviews help ensure that third-party AI services continue to meet internal requirements.

    AI TRiSM: Example Workflow 

    Consider an ecommerce company that uses an AI customer service agent to answer support requests and perform actions such as checking orders, issuing refunds, and updating customer accounts. A customer asks the agent to refund a recent purchase to a different payment method than the one originally used.

    This single request can activate several AI TRiSM controls before the agent is allowed to take action.

    1. Discovery Identifies the AI System and Its Approved Use

    The AI TRiSM system first checks the organization’s AI inventory to identify the customer service agent, the model it uses, the systems it can access, and its approved use cases.

    It retrieves information such as the agent’s risk classification, authorized tools, maximum refund amount, and rules governing access to customer and payment data. This establishes what normal and permitted behavior should look like for the agent.

    2. Information Governance Checks Data Access

    Next, information governance controls determine whether the agent is authorized to access the customer’s order, account information, and payment details.

    The system verifies the customer’s identity and ensures the agent only receives the information necessary to process the request. Sensitive data such as full payment card details can be masked or excluded from the model’s context entirely.

    If the requested information falls outside the agent’s permitted access, the workflow can stop or be escalated to a human representative.

    3. Runtime Inspection Evaluates the Request and Proposed Action

    Once the data checks pass, runtime controls evaluate the customer’s request, the context retrieved from backend systems, the model’s response, and any tool calls the agent attempts to make.

    For example, one policy engine might detect that the customer requested a refund to a payment method that does not match the original transaction. Another could determine whether the refund amount exceeds the agent’s approved limit, while additional controls check for prompt injection, suspicious account activity, sensitive data leakage, or unauthorized tool use.

    Runtime inspection also verifies that the agent’s proposed actions remain within its approved scope before allowing it to call payment or account-management systems.

    4. Risk Signals Are Combined Into an Enforcement Decision

    The results of these checks are combined into an overall risk score that determines what happens next.

    • High risk: The refund action is blocked, such as when the agent attempts to transfer money to an unauthorized payment method.
    • Medium risk: The request is paused and routed to a fraud, security, or customer service representative for review.
    • Low risk: The agent is permitted to complete the refund and confirm the result to the customer.

    This allows multiple governance, security, and data signals to influence a single enforcement decision rather than relying on one isolated rule.

    5. The Interaction Is Logged and Fed Back Into Governance

    Finally, the system records the customer request, data accessed, tools invoked, risk signals, enforcement decision, and any human escalation.

    Governance and security teams can use these records to identify recurring problems and adjust controls. For example, repeated attempts to redirect refunds to different payment methods could lead the organization to tighten refund policies or require human approval for that action.

    The result is a continuous AI TRiSM cycle: governance defines what the agent is allowed to do, information governance defines what data it can use, runtime controls evaluate what it actually does, and feedback helps improve those controls over time.

    How to Get Started With AI TRiSM Implementation

    Implementing AI TRiSM is usually an incremental process rather than a single technology deployment. Organizations can start by identifying where AI is already being used, classifying the associated risks, and then applying governance, data, runtime, and infrastructure controls based on business impact. The goal is to create a repeatable process for evaluating AI systems before deployment and continuously monitoring them once they are in use.

    Build an Inventory of AI Systems and Use Cases

    Start by identifying the AI models, applications, agents, copilots, and third-party AI services used across the organization. The inventory should include both officially approved systems and shadow AI that employees may be using without formal oversight, such as public generative AI tools or unsanctioned SaaS applications with embedded AI features.

    For each system, document its owner, business purpose, underlying model, deployment environment, users, data sources, integrations, connected tools, and level of autonomy. Organizations should also record whether the system can generate content, access internal information, make recommendations, or take actions in external systems. This inventory becomes the foundation for assessing risk and determining which AI systems need the strongest controls.

    Classify AI Systems by Risk

    Not every AI workload requires the same level of oversight. Organizations should classify AI systems according to factors such as data sensitivity, business impact, regulatory exposure, user reach, model autonomy, and the consequences of incorrect or unauthorized actions.

    For example, an internal writing assistant that only works with public information may represent relatively low risk, while an AI agent that can modify financial records, approve transactions, or access customer data may require strict controls and human oversight. Risk classifications can be used to determine testing requirements, approval processes, access restrictions, monitoring intensity, and escalation procedures.

    Organizations should also reassess classifications when an AI system changes. Adding new tools, connecting additional data sources, expanding the user base, or increasing an agent’s autonomy can significantly change its risk profile.

    Define AI Governance Policies and Ownership

    Establish clear policies that define how AI systems can be developed, purchased, deployed, and used across the organization. These policies should address acceptable AI use, model approval, risk assessments, testing requirements, data handling, human oversight, access controls, documentation, and change management.

    Governance should also define who is responsible for different types of AI risk. Security teams may focus on technical threats, while legal and compliance teams address regulatory requirements and business units remain accountable for the systems they deploy.

    Clear ownership prevents situations where AI systems operate without anyone being responsible for their behavior or associated risks. Organizations can also establish review boards or approval workflows for higher-risk AI applications so security, compliance, data, and business stakeholders can evaluate them before deployment.

    Establish Information Governance and Data Controls

    Identify the data that AI systems are permitted to access and classify sensitive information such as personal data, credentials, intellectual property, financial information, source code, and regulated records. Organizations should understand not only the data used to train or fine-tune models, but also the information supplied through prompts, retrieval systems, plugins, and connected enterprise applications.

    Apply access controls, data loss prevention, retention policies, encryption, and data minimization so models and agents receive only the information necessary for their tasks. Sensitive information can also be masked, redacted, or tokenized before it reaches an AI model.

    Data governance policies should extend to third-party AI services as well. Organizations need to understand whether external providers retain prompts, use submitted information for training, or transfer data across jurisdictions before allowing sensitive information to be processed.

    Implement Runtime Monitoring and Policy Enforcement

    Deploy controls that inspect AI activity while systems are operating, rather than relying solely on predeployment assessments. Runtime monitoring can evaluate prompts, model outputs, retrieved content, API requests, tool calls, and agent actions for security threats, policy violations, and abnormal behavior.

    For example, runtime controls can detect prompt injection attempts, attempts to expose sensitive information, prohibited content, unusual tool usage, or an AI agent attempting actions outside its approved scope. Policies can then determine whether to allow the interaction, sanitize content, mask data, block the request, or require human approval.

    Continuous runtime visibility is particularly important for generative AI and autonomous agents because their behavior can vary depending on context, user input, retrieved information, and connected tools. Monitoring actual behavior provides insight into risks that may not appear during controlled testing.

    Secure the AI Infrastructure and Integrations

    Protect the infrastructure supporting AI systems, including APIs, model endpoints, vector databases, MCP servers, plugins, cloud environments, orchestration frameworks, and development pipelines. Each component can introduce additional attack paths or permissions that attackers may attempt to exploit.

    Use strong authentication, least-privilege access, network segmentation, workload isolation, API security, secrets management, encryption, and detailed logging. Credentials used by agents or AI applications should be scoped narrowly and rotated regularly rather than giving models broad access to enterprise systems.

    Organizations should also assess third-party models, libraries, APIs, plugins, and MCP servers before connecting them to sensitive environments. A compromised integration or overly permissive tool can undermine otherwise strong AI governance and expose data or systems to unauthorized activity.

    Continuously Test, Monitor, and Improve Controls

    AI TRiSM should continue throughout the entire lifecycle of an AI system. Organizations should monitor model behavior, policy violations, security events, performance drift, data changes, access patterns, and modifications to connected systems after deployment.

    Regular red teaming, adversarial testing, access reviews, model evaluations, and risk reassessments can reveal weaknesses that were not apparent during initial testing. Teams should also monitor emerging AI threats and update detection rules and governance policies as attack techniques evolve.

    The results of monitoring and testing should feed back into the broader AI TRiSM program. For example, repeated policy violations may indicate that access controls need tightening, while frequent human escalations may suggest that an agent’s permissions or decision thresholds should be adjusted. This feedback loop helps AI governance and security controls evolve alongside the systems they protect.

    Securing AI Agents and Enforcing AI TRiSM Controls with Exabeam Agent Behavior Analytics

    Exabeam Agent Behavior Analytics (ABA) extends behavioral analytics to AI agents and autonomous workflows, giving security teams visibility into AI agent activity across the digital workforce. The platform applies behavioral intelligence to establish baseline normal activity and surface subtle anomalies across agent operations, detecting misuse, drift, abnormal tool usage, risky access, and actions outside defined roles. To support compliance with AI TRiSM frameworks, Exabeam contributes to open-source initiatives like Praxen, which provides pre-deployment Agent Behavior Verification (ABV) to align agent boundaries with organizational policies, and Observra, which provides standardized runtime telemetry to ensure AI systems operate within defined guardrails. Activity is correlated with users, entities, applications, and security events, helping teams identify emerging agent risk and investigate it before it impacts operations.

    Key capabilities of Exabeam Agent Behavior Analytics:

    • Behavioral detections for AI agents: Detects autonomous installation and execution, tracks tool usage and MCP activity, and identifies agent activity that falls outside expected roles.
    • Broad AI platform coverage: Analyzes activity from Claude Enterprise, ChatGPT, Google Gemini, Microsoft Copilot, GitHub Copilot, and custom agents that share open telemetry.
    • Open agent telemetry: Uses Observra to standardize agent-native telemetry and Exabeam Agent Sensor to collect and normalize activity across models, agents, and runtime environments.
    • AI security posture management: Outcomes Navigator maps ABA coverage to the OWASP Agentic Top 10, identifies gaps in data and detections, and provides prioritized recommendations to measure and improve readiness.
    • Automated investigation of agent activity: Machine-built timelines correlate agent actions with related user, entity, and application activity, while Exabeam Nova AI summarizes threats, explains behavioral context, and recommends next steps with analysts remaining in control.
    • Customizable detection rules: Teams can review, clone, tune, and customize ABA detections, and use Nova AI rule creator to generate behavioral models and correlation rules from natural language.
    • Flexible deployment: Add ABA to an existing SIEM with New-Scale Analytics, or deploy New-Scale Fusion for integrated detection, investigation, and response across human and non-human identities.
    • Open-source verification and telemetry: Praxen enables pre-deployment verification by comparing declared behavioral policies with actual implementation evidence. Meanwhile, Observra provides a standardized open-source telemetry library that captures and normalizes agent activity across major frameworks to feed directly into security monitoring.

    Learn more about Exabeam Agent Behavior Analytics

    Learn More About Exabeam

    Learn about the Exabeam platform and expand your knowledge of information security with our collection of white papers, podcasts, webinars, and more.

    • Webinar

      The Insider Risk Program Just Got Bigger: How Digital Workers Are Changing Identity, Behavior, Ownership, and...

    • Blog

      Every AI Agent Creates Security Debt. Most Organizations Aren’t Tracking It.

    • White Paper

      Detection Strategies for the Agentic Enterprise

    • Blog

      What Is the Difference Between Exabeam and CrowdStrike for SIEM and Security Operations?

    • Show More
    Add to Preferred Sources