Leakage rate is the share of AI outputs that are blocked, redacted, or otherwise prevented from being shown because they would violate policy. It is one of the clearest indicators that an AI system is crossing data boundaries or that the underlying classification and entitlement model needs tuning.
Expanded Definition
Leakage rate measures how often an AI system’s outputs are suppressed, redacted, or blocked because policy engines detect disclosure risk. In NHI and agentic AI environments, it is less about model quality and more about boundary control: what the system is allowed to reveal, to whom, and under what context. Definitions vary across vendors, but the operational meaning is consistent enough for governance: a rising leakage rate usually signals either overexposure of sensitive context or overly broad prompting, retrieval, or tool access. It should be read alongside entitlement design, data classification, and guardrail tuning rather than treated as a standalone model score. For AI governance context, the NIST AI Risk Management Framework is useful because it frames harmful output, privacy, and governance as lifecycle risks rather than isolated incidents. The most common misapplication is treating leakage rate as proof of model failure when the real issue is weak classification, incomplete redaction rules, or excessive context injection.
Examples and Use Cases
Implementing leakage rate monitoring rigorously often introduces friction, because tighter controls can reduce answer completeness and increase false positives, requiring organisations to balance usability against exposure risk.
- A support chatbot retrieves customer records and the policy layer blocks account numbers before they are shown to the user.
- An internal agent generates a draft incident summary, but redaction rules suppress API keys, token fragments, and embedded credentials.
- A retrieval-augmented generation workflow returns a document passage that contains privileged data, and the response is withheld until access is verified.
- A security team compares leakage rate spikes against changes in prompt templates or tool permissions after reviewing lessons from the Ultimate Guide to NHIs — Why NHI Security Matters Now.
- Policy tuning is informed by broader secret-sprawl lessons from the Guide to the Secret Sprawl Challenge, especially when outputs reveal credentials copied into code or chat context.
For implementation guidance, teams often compare blocked-output trends with the disclosure risks highlighted in NIST AI Risk Management Framework and adjust their guardrails accordingly.
Why It Matters for Security Teams
Leakage rate matters because it exposes how close an AI system is to crossing a trust boundary. A low rate can indicate strong controls, but it can also hide underblocking if sensitive data is not being detected. A high rate can mean the system is being asked to operate on overly broad context, or that the underlying classification model is too aggressive and degrading utility. For NHI and agentic AI security, this is especially important because service accounts, tokens, and delegated tool permissions often create the exact conditions where private context becomes machine-readable and then machine-reissued.
NHIMG research shows why this cannot be ignored: 52 NHI Breaches Analysis and the Ultimate Guide to NHIs — Why NHI Security Matters Now both document how secrets exposure and excessive privileges amplify downstream damage. NHI Mgmt Group research also reports that 79% of organisations have experienced secrets leaks, and 77% of those incidents caused tangible damage. The practical lesson is that leakage rate is not just a model metric, it is a governance signal that the access model, retrieval scope, or secret handling path is failing.
Organisations typically encounter leakage rate as a business problem only after a blocked prompt, incident review, or audit finding reveals that sensitive data was already entering the AI workflow, at which point the term becomes operationally unavoidable to address.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and OWASP Non-Human Identity Top 10 address the attack and risk surface, while NIST AI RMF, NIST AI 600-1 and NIST CSF 2.0 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST AI RMF | Frames AI output harms, privacy risk, and governance across the AI lifecycle. | |
| NIST AI 600-1 | Profiles GenAI risks including harmful disclosure, privacy, and guardrail behavior. | |
| OWASP Agentic AI Top 10 | Covers agentic AI risks where tool access and prompt flow can disclose sensitive data. | |
| OWASP Non-Human Identity Top 10 | NHI-02 | Secret exposure and improper handling are core NHI governance concerns. |
| NIST CSF 2.0 | PR.DS | Data security controls govern protection of information as it moves through systems. |
Treat leakage rate as a lifecycle risk signal and tune controls where outputs expose sensitive context.
Related resources from NHI Mgmt Group
- How can organisations reduce secret leakage in ServiceNow at scale?
- What is the difference between source control leakage and SharePoint secret exposure?
- How should security teams reduce secrets leakage without slowing developers down?
- Why do AI agents create more leakage risk than traditional applications?
Deepen Your Knowledge
Reviewed and updated by the NHIMG editorial team on July 9, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org
