AI governance platforms are exposing gaps in runtime AI risk

At a glance

What this is: NHI Mgmt Group's analysis, drawing on WitnessAI's blog post: Policy-only governance is no longer enough when shadow AI, sensitive data exposure, and agentic workflows all create runtime risk across five AI governance platforms.

Why it matters: IAM, security, and governance teams need a way to separate visibility add-ons from runtime control layers, because AI governance now intersects with identity, data handling, and machine-to-tool execution.

Context

AI governance platforms now sit at the intersection of discovery, runtime enforcement, and risk management. The problem is no longer limited to policy documentation or audit readiness, because employees, production models, and AI agents can all touch sensitive data and external tools before controls see them.

The governance gap is visible in three places at once: shadow AI that security teams cannot inventory, model interactions that need runtime inspection, and agentic workflows that can initiate API calls or database queries without a human decision point. That makes AI governance a control problem as much as a policy problem.

For identity and access teams, the category matters because it extends beyond conventional GRC into runtime authorisation, data handling, and tool-using AI behaviour. The article is essentially about matching platform architecture to the actual AI footprint, not to an abstract compliance model.

Key questions

Q: What breaks when AI governance only documents policy and does not inspect runtime activity?

A: Policy-only governance leaves a blind spot at the point where AI systems actually process prompts, route data, and trigger actions. That means shadow AI, unsafe outputs, and delegated tool use can all proceed without intervention. The failure is not lack of documentation, but lack of enforcement where the risk occurs.

Q: Why does shadow AI create a governance risk even when employees mean well?

A: Because intent does not remove exposure. Employees may adopt AI to work faster, but unapproved tools can still receive proprietary content, bypass review expectations, and create compliance gaps. The risk comes from unmanaged data flow and lack of oversight, not from employee malice.

Q: What are the signs that an AI governance platform is too limited for enterprise use?

A: Common warning signs include browser-only visibility, a narrow catalog of supported tools, weak coverage of native applications, and no clear runtime control over prompts, responses, or action execution. Those limits usually mean the platform can report on AI use but cannot meaningfully govern it.

Q: Should organisations extend zero trust or adopt a dedicated AI governance platform?

A: That depends on the AI footprint and the gap you are trying to close. If the issue is visibility for existing users and apps, an existing zero-trust or SSE stack may be enough. If the problem is shadow AI, runtime policy, and agentic actions, a dedicated AI governance layer is usually the cleaner fit.

Technical breakdown

Why policy-only AI governance misses runtime behaviour

Policy documentation can record intent, but it cannot stop an AI tool from sending prompts, forwarding context, or invoking external services at execution time. Runtime governance means controls inspect the interaction as it happens, not after the fact. That distinction matters when the system in question can query databases, call APIs, or generate responses that expose sensitive information. Traditional governance workflows were built for approval and evidence collection, not active interception of AI behaviour. In practice, the control boundary moves from paperwork to enforcement points that see the live session, the data being passed, and the action being attempted.

Practical implication: validate whether a platform can enforce policy during the AI session, not just document policy after deployment.

Shadow AI discovery needs visibility beyond the browser

Shadow AI is often missed because it does not always appear in the same place as web traffic. Employees use personal accounts, native desktop apps, embedded copilots, IDE plugins, and model access paths that never pass through a browser-only control layer. Discovery therefore depends on seeing the real enterprise AI footprint, including sanctioned and unsanctioned usage, rather than relying on a tracked catalog or a single egress point. If the platform cannot observe native applications and embedded workflows, its inventory will always undercount the actual attack surface and data exposure surface.

Practical implication: compare the platform’s coverage of browser, native, and embedded AI usage before treating its inventory as complete.

Agentic workflows create a new authorisation problem

Agentic AI changes governance because the system can act on behalf of a user by selecting tools, making API calls, and chaining actions across multiple steps. That is not just a data leakage issue. It is an authorisation problem, because the agent may move from permissible intent to operational execution inside one session. MCP-connected tools and other orchestration layers make that behaviour more consequential by extending the agent’s reach into external systems. Once the workflow can initiate actions without a human in the loop, the governance model has to account for tool use, action scope, and oversight at runtime.

Practical implication: require explicit runtime authorisation controls for any AI workflow that can call tools or execute multi-step actions.

Breaches seen in the wild

• Samsung ChatGPT leak 2023: Samsung staff pasted chip source code and meeting notes into ChatGPT weeks after it was allowed, leading Samsung to restrict generative AI tools.
• OmniGPT breach claim 2025: A hacker claims to have leaked 34 million OmniGPT AI chat messages holding users' API keys and credentials; OmniGPT has not confirmed it.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

AI governance has moved from documentation to runtime control. The category now has to answer whether an organisation can see AI use as it happens and intervene before sensitive data or unsafe actions leave the boundary. Policy artifacts alone do not address that problem because they do not govern live execution. The practical conclusion is that governance platforms must be judged by enforcement depth, not by their reporting surface.

Shadow AI is a discovery problem before it is a policy problem. If a platform only sees browser traffic or a narrow application catalog, it will miss the systems employees actually use. That creates a false sense of coverage and leaves unmanaged AI channels intact. The practitioner lesson is that inventory fidelity must be tested against the full enterprise footprint, including native apps and embedded AI.

Agentic AI expands the control surface into authorisation, not just content moderation. When an AI system can call tools, query databases, and chain actions, the issue is no longer only what it says. The system is operating with delegated execution authority, which means control design has to account for who can trigger actions, what those actions can reach, and how runtime oversight is applied. The implication is that AI governance and identity governance are converging around execution rights.

Runtime AI risk management is becoming the defining governance gap. The article’s core tension is not vendor selection, but whether organisations are trying to manage AI with static GRC models that were never built to observe live behaviour. A named concept here is runtime AI risk gap: the distance between documented policy and enforceable control at the point of model, agent, or tool execution. Practitioners should treat that gap as a design fault, not an adoption phase.

Dedicated AI governance platforms are pulling security decisions down to the control plane. The article shows that visibility, enforcement, and auditability are being consolidated into runtime architecture rather than left in separate compliance workflows. That shift matters because it changes who owns AI risk in practice: security, infrastructure, and identity teams will all influence the control model. The field implication is that AI governance is becoming an operational security discipline, not a reporting exercise.

From our research library:

• 52% of respondents see AI security decision-making power shifting toward platform and infrastructure teams rather than the executive suite, according to the 2026 Infrastructure Identity Survey.
• Read next: AI Agent Observability, Audit and Incident Response Guide

What this signals

Runtime AI risk gap: enterprises should expect governance programmes to be judged by live enforcement, not by policy documentation or audit artefacts. That means the next control conversation is about what the platform can see and stop at execution time, especially when AI reaches beyond browser-based use.

The purchase decision is increasingly architectural. Teams need to separate visibility add-ons from controls that can inspect prompts, constrain tools, and govern agentic workflows, because those are different operating models with different failure modes.

The 52% figure showing decision-making power shifting toward platform and infrastructure teams suggests AI governance is moving out of traditional policy ownership and into operational control. That shift should change who is accountable for discovery, runtime enforcement, and evidence generation in the AI stack, according to the 2026 Infrastructure Identity Survey.

For practitioners

• Validate runtime enforcement before buying visibility Test whether the platform can inspect and control prompts, responses, and actions while the session is live. If it only produces reports after the fact, it is not closing the runtime risk gap.
• Map the full AI footprint across channels Check whether discovery covers browser use, native desktop apps, embedded copilots, IDEs, and approved and unapproved AI services. Treat incomplete coverage as a blind spot, not a reporting limitation.
• Separate policy tooling from execution controls Use governance records for audit evidence, but require a separate control layer for data protection, tool access, and runtime decision points in AI workflows.
• Test agentic workflows for delegated action scope Review any AI system that can call tools or query data as a delegated executor, and verify what actions it can initiate without human approval.

Key takeaways

• AI governance failures now come from runtime gaps as much as from missing policies, because live model and agent behaviour can outrun documentation.
• Discovery, enforcement, and audit need to work together if organisations want coverage across shadow AI, sanctioned AI, and agentic workflows.
• The practical test is simple: if a platform cannot see the AI activity and control it as it happens, it is not enough for enterprise governance.

Key terms

• Runtime AI Risk Management: The ongoing operational discipline of identifying, controlling, and evidencing AI risk while the system is in production. It focuses on live prompts, responses, policy outcomes, and audit trails rather than static policy documents or one-time approval decisions.
• Shadow AI: AI agents, copilots, or connected tools operating without full visibility or governance from security teams. Shadow AI becomes an identity problem when those systems authenticate with unmanaged tokens, service accounts, or OAuth apps that can reach production resources.
• Agentic workflow: An agentic workflow is a sequence of tasks executed by an AI agent with some level of tool access and decision authority. In security terms, the workflow matters because it can span multiple systems, identities, and permissions, which makes attribution and revocation harder than with ordinary automation.
• Runtime Enforcement: Runtime enforcement is the practice of blocking malicious behaviour while software is running, rather than only detecting it after the fact. It monitors process activity, network actions, and privilege changes so a live attack can be interrupted at the point of execution.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are responsible for identity security strategy or NHI governance in your organisation, it is worth exploring.