Gartner's AI trust, risk, and security management (TRiSM) framework describes what AI systems need to be to stay trustworthy in production. The framework states that these systems need to be explainable, monitored, secure, and careful with the data they touch. The framework does not set any of that. Security and IT teams still have to turn each pillar into controls with owners, evidence, and a safe way to fail.
What is AI trust, risk, and security management?
AI TRiSM is the layer of technical controls that keeps AI systems trustworthy, secure, and compliant while they run. AI TRiSM has four pillars:
Explainability and model monitoring
ModelOps
AI application security
Gartner's public AI readiness page lists AI TRiSM technology functions including AI governance, AI runtime inspection and enforcement, and information governance.
AI governance is the set of decisions and accountabilities, such as who approves a use case, what risk is acceptable, and who answers when a system fails. TRiSM provides technical controls to put those decisions into operation.
Which controls to set for explainability and model monitoring
Models degrade even when nobody changes them. Input data and real-world conditions shift after launch, so a model that tested well can lose quality without any change to its code. Monitoring catches model decay before a customer or auditor does.
Output records: For each consequential output, store the input, model and prompt version, retrieved context, and any rationale the system produced. A reviewer should be able to reconstruct why the system answered as it did.
Drift and anomaly detection: Baseline input distributions and output quality at launch, then alert when either moves past a set threshold. For large language model (LLM) outputs, track grounded-response and unsupported-output rates alongside accuracy.
Continuous evaluation: Run offline test sets before each release and sampled evaluation in production after it. Keep human review on high-risk paths. LangChain's 2025 survey of 1,340 practitioners (LangChain sells evaluation tooling) found that 89% had observability in place, but only 52% ran offline evaluations and 37% ran online evaluations.
Decision rules: Decide in advance who chooses between accepting, retraining, rolling back, or withdrawing a model when a threshold is breached.
Model owners in data science or AI engineering typically own these monitors, with anomaly alerts routed to security.
Which controls to set for ModelOps
ModelOps (model operations) turns "we approved this model" into evidence an auditor can check. Approval evidence often breaks down under pressure. In an EY survey reported by Dark Reading, 47% of respondents said their organization had skipped its AI governance process for urgent deployments, even though 98% had formal policies.
Five lifecycle controls carry over from software delivery, extended to models, prompts, tools, and data.
Versioning: Version the model, prompt templates, tool configurations, and evaluation data together. A model version alone cannot explain an agent's behavior.
Pre-deployment testing: Run each candidate against the evaluation suite and adversarial cases before promotion.
Named approver: Require a named person, separate from the requester, to approve promotion to production.
Rollback: Test rollback before it is needed. When a Replit agent deleted a production database during a code freeze, CEO Amjad Masad said the company responded by separating development and production databases and improving rollback.
Audit trail: Record who changed what, who approved it, and which tests ran.
Rollback depends on people knowing the process exists. ISACA's 2026 AI Pulse Poll found that 39% of respondents did not know whether their organization had a documented process for shutting down or overriding AI systems.
In Tines 3B, the complete environment for enterprise AI, teams test changes on an isolated branch before pushing them live, and every commit is kept for rollback to a known-good version.
Which controls to set for AI application security
Generative AI (GenAI) and agent attacks often arrive through content the model reads, such as emails, documents, and retrieved pages. Gartner predicts that by 2029, over half of successful cyberattacks on AI agents will exploit access control weaknesses and prompt injections.
Four threats stand out, and each maps to a specific control:
Prompt injection: EchoLeak (CVE-2025-32711) let a single crafted email cause Microsoft 365 Copilot to exfiltrate internal files with no user interaction, and Microsoft patched it server-side. The Open Worldwide Application Security Project (OWASP) says it is unclear whether fool-proof prevention methods exist. Set up input and output filtering, mark untrusted content as untrusted, validate output formats with deterministic code, and pair these with least-privilege tool access so a successful injection has little it can reach.
Data poisoning: Research from Anthropic, the UK AI Security Institute, and the Alan Turing Institute found that 250 malicious documents could backdoor models of very different sizes, using simple, low-stakes backdoors. Track data provenance, vet data sources, sandbox exposure to unverified data, and check outputs against trusted references.
Oversharing: OWASP warns that system prompt restrictions on what data an LLM returns can be bypassed through prompt injection. Enforce least privilege at the data layer, and redact or tokenize sensitive fields before the model sees them.
Excessive agency: OWASP traces this to excessive functionality, permissions, and autonomy. Give agents only the tools a task needs, scope credentials to that task, require human approval for high-impact actions, and enforce authorization in downstream systems instead of trusting the model's judgment.
These controls share one runtime layer of guardrails, sandboxing, and least-privilege credentials. In Tines, each workflow step runs in its own isolated sandbox, and a credential proxy (a service that holds secrets and adds them to outbound requests) injects secrets at runtime, so workflow code and the model never see them.
Which controls to set for privacy and data protection
AI tools see whatever employees and agents feed them, so privacy controls decide which data reaches a model and where it ends up. In a Gartner survey of 175 employees, over 57% used personal GenAI accounts for work and 33% admitted entering sensitive information into unapproved tools. IBM's 2025 Cost of a Data Breach report found that organizations with high levels of shadow AI paid $4.74 million per breach, $670,000 more than those with low or no shadow AI.
Five controls address these risks.
Classification: Label which data classes may enter prompts, retrieval indexes, and fine-tuning sets, and block the rest at the pipeline.
Leakage controls: Inspect prompts and outputs for regulated data, and redact or tokenize it before processing.
Access scoping: Make retrieval respect source-system permissions, and have agents act in the requesting user's context with minimum scope.
Secrets handling: Keep API keys out of prompts and code (the credential proxy described above handles this). Inject them at runtime and rotate them.
Sanctioned alternatives: Give employees an approved tool they want to use. Blanket bans tend to remove visibility while the behavior continues.
Three places deserve the same scrutiny as prompts. First, retrieval indexes built from internal documents can expose that content to anyone who can query them. Second, agents with write access to systems of record can change data that other teams rely on. Third, the logs that store full prompts also hold sensitive data, so prompt logs are a sensitive data store in their own right and need the same retention and access rules.
How AI TRiSM maps to the EU AI Act, NIST AI RMF, and ISO/IEC 42001
All three frameworks ask for evidence that controls exist and work. The table below shows which TRiSM pillar supplies that evidence for each framework.
Evidence such as logs, approvals, and test results is easiest to produce when the workflow that enforces a control also records it. Teams building a release approval gate that records who signed off and which tests ran can start from the workflow examples gallery, which collects IT and security starting points.
How to implement AI TRiSM
Each control works only when a named person answers for it, and responsibility for AI governance is spread across several functions. The International Association of Privacy Professionals (IAPP) found that privacy teams and legal and compliance teams each hold primary AI governance responsibility at 22% of organizations, IT at 17%, and security at 5%.
Each of the seven steps below gets an explicit owner:
Inventory every AI system, model, and agent. The chief information officer (CIO) owns the register, and security runs discovery for shadow AI. A Cloud Security Alliance survey found that only 21% of organizations keep a real-time registry of their agents.
Classify the data each system touches. Data governance owns this step, with privacy.
Risk-tier each system. Risk and compliance owns tiering, with legal. Gartner warns that applying uniform governance to all AI agents can lead to enterprise AI agent failure.
Gate releases. The model owner proposes, and a named approver from risk or security signs off.
Inspect at runtime. The chief information security officer (CISO) owns guardrails, sandboxing, and anomaly alerts.
Require approval for high-impact actions. The business owner of the affected system approves.
Log and review. Security keeps the logs, and internal audit reviews them on a fixed schedule.
The standards reviewed here name measurement categories but set no numeric thresholds, so the measures below are recommendations rather than requirements. Four measures show whether the program works:
Inventory completeness: Count models found by discovery sweeps that were missing from the register, against a target of a complete inventory.
Validation coverage: Track the share of high-tier systems with a current independent validation, targeting 100%.
Drift response time: Measure the median time from a drift breach to a decision to accept, retrain, or withdraw.
Policy exception rate: Track exceptions to the standard process to identify where governance is being bypassed.
Tracking these measures depends on the same logs, approvals, and runtime records that the controls above produce.
Putting AI TRiSM into practice with Tines 3B
Security and IT teams use Tines 3B as the governed environment where they run AI agents and AI-generated workflows under controls they set themselves. Teams define those controls, such as inventory sweeps, approval gates, and runtime checks, to fit their own policies. Audit logs record every user and system action, alongside the step-level sandboxing described earlier.
Teams read those audit logs as step seven's review evidence, and role-based access control (RBAC), scoped per space, connector, and network, limits who can reach which systems and credentials. Paid plans let teams bring their own model, so model choice stays a governance decision.
The pillars hold up only once each one becomes a control with a named owner, evidence of operation, and a defined way to fail safely. In Tines, teams run those controls with release gates, runtime credentials, and rollback. A practical first step is to pick one high-tier AI use case, put its release gate and runtime credentials under these controls, and test that the rollback works. Teams can start in Tines 3B Explore Edition, a free version with limits that includes AI, single sign-on (SSO), and unlimited integrations.
