AI Governance Glossary

An NCUA letter to credit unions (2007, issued with Supervisory Letter 07-01) that sets the still-operative framework for evaluating third-party relationships. Its central idea is proportionality: due diligence should match the risk, complexity, and criticality of the vendor, with "reasonable alternative procedures" for non-complex, non-core vendors. NCUA never adopted the 2023 interagency third-party guidance, so for credit unions this is the operative standard.

An open protocol, introduced by Google in 2025 and later contributed to the Linux Foundation, that lets independent AI agents discover each other's capabilities, exchange information, and coordinate tasks across vendors and frameworks. In plain terms: a standard for your vendors' AIs to talk to each other's AIs. That raises an oversight question almost no one is asking yet — who is accountable when the agents on both sides of a transaction are autonomous.

AI that doesn't just generate text but takes actions toward a goal — calling tools, reading and writing data, and making multi-step decisions with limited human intervention. It's the category that turns "an AI wrote a draft" into "an AI clicked submit." US bank regulators explicitly carved generative and agentic AI out of model-risk scope in SR 26-2 — a gap, not a green light.

A software program that uses an AI model to perform tasks on a user's behalf — reading data, calling APIs, and acting across systems, often without a human approving each step. The compliance problem it creates is identity: when an agent takes an action, "who did this?" can have no clean answer.

Applying the third-party risk management discipline you already run to AI tools and AI-enabled vendors specifically — what the tool can access, what it does with the data, whether it's retained or used for training, and whether the vendor can prove its controls. As of mid-2026, no incumbent sells one at small-institution price points.

Marketing a product or service as "AI-powered" when it isn't, or overstating what the AI actually does. It matters to RIAs because the SEC has treated it as a real disclosure risk — bringing enforcement actions against advisers for overstating AI use and naming AI claims as an area of exam focus. Your AI-related statements should be consistent with your Form ADV and with what your tools actually do.

A security tool that sits between users and cloud services to monitor and enforce policy on the traffic between them. It's worth knowing because it's what people assume shadow-AI discovery means — network interception. NobleCloak deliberately does not work this way: our discovery is an OAuth grant inventory, not traffic interception.

Ongoing, usually automated observation of a risk over time — as opposed to a snapshot. It answers "is this still true today?" rather than "was this true on the scan date?" Both have a place; the honest question is which one your examiner is actually asking for.

How long a vendor keeps your data after it's done processing it — and whether you can control or delete it. For AI tools this is a first-order due-diligence question: a transcription tool that retains recordings indefinitely is a very different risk than one that discards them on delivery.

NobleCloak's deliverable: a point-in-time AI vendor-risk assessment of your organization's AI exposure — every AI tool your people authorized, the data each can reach, who's using it, and how it maps to your regulatory frameworks. It reads your SSO- and OAuth-discoverable AI surface and states its coverage boundary on every report.

Technology that detects and blocks sensitive data from leaving an organization — email, uploads, downloads. It's a useful control, but it is not vendor diligence: DLP tells you data moved, not whether the AI vendor on the other end is safe to trust with it.

The European Union's AI regulation, which entered into force on August 1, 2024 and phases in through 2026 on a risk-based model (prohibited, high-risk, limited-risk, minimal-risk uses). It is not a US rule and does not bind a US credit union or RIA directly — but its requirements are increasingly written into the contracts and representations of the global vendors you buy from.

The organized collection of documentation that proves your controls to an examiner — policies, vendor assessments, findings, and the mapping that ties them to a framework. The whole point of an AI vendor-risk program is to produce one you could hand over without a scramble.

Evidence structured to survive an exam: in the examiner's vocabulary, mapped to the frameworks they cite, dated, and exportable — not a folder of screenshots you have to explain. "Examiner-ready" is a claim about usability under questioning, not just completeness.

The disclosure form investment advisers file with the SEC or state regulators, describing their business, practices, and conflicts. Your statements about AI — in the ADV and in marketing — need to be consistent with what your firm and its vendors actually do, or you risk an AI washing finding.

A May 2025 report from the US Government Accountability Office finding that NCUA lacks the authority to examine third-party technology vendors — even for AI-driven services. GAO has recommended a legislative fix since 2015; Congress has not enacted one. The practical effect: for credit unions, the vendor-diligence burden lands structurally on you, with no regulator doing it for you.

The 1999 federal law requiring financial institutions to protect the security and confidentiality of customer information. Reg S-P is the SEC's implementation of its safeguards and privacy provisions — so when people say "Reg S-P," GLBA is the statute underneath it.

Guidance issued by the Federal Reserve, FDIC, and OCC in 2023 on managing risk from third-party relationships (sometimes cited as "2023-17"). It applies to the banks those agencies supervise. NCUA did not adopt it — credit unions remain under 07-CU-13 — so be careful not to cite it as a credit-union standard.

An open standard, introduced by Anthropic in November 2024 and later contributed to a Linux Foundation effort, that lets AI models connect to external data sources and tools through "MCP servers." Think of an MCP server as a connector that hands an AI a key to a system — your files, your CRM, your calendar. That connector is, in every sense that matters to your program, a third-party vendor with data access.

Whether a vendor uses the data you put into its AI — your prompts, documents, client information — to train or improve its models. It is one of the sharpest due-diligence questions you can ask, because a "yes" (or a vague non-answer) means your customers' data may live inside a model you don't control. NobleCloak, for its part, does not train on your data.

The NIST AI Risk Management Framework 1.0, released January 26, 2023 — a voluntary framework organized around four functions: Govern, Map, Measure, Manage. It isn't a regulation and imposes no mandate, but it's the common language examiners and vendors increasingly reach for, which makes it an excellent crosswalk for your own findings.

An identity that belongs to a machine, service, or agent rather than a person — an API key, a service account, an OAuth token. As AI agents proliferate, non-human identities multiply, and each one is a credential that can reach data and take action. They're often the least-governed part of a stack.

The permission a user gives a third-party app to access their account data — calendar, email, files — without handing over their password. It's the mechanism behind most "connect your account" buttons, and it's the persistent key an AI tool holds long after the click. Reading OAuth grants across your identity provider is how you map your real AI surface.

An evaluation of risk as of a specific date — a snapshot. It's the honest description of what a Discover scan is, and it's genuinely sufficient for a great deal of exam evidence. Where it isn't enough is drift: what changed after the scan.

An attack in which hidden or malicious instructions — buried in a document, a web page, or an email an AI reads — trick the AI into doing something it shouldn't, like leaking data or taking an unauthorized action. It's the AI-era version of a phishing payload, aimed at the model instead of the user.

Access that can view but never change or act. The Discover scan runs on read-only access to your identity provider — it can see what AI tools exist and what they can reach, and it can do nothing else. Connecting an AI to govern it is a separate, larger consent you grant knowingly; read-only means read-only.

SEC Regulation S-P, which implements GLBA's safeguards and privacy rules for the firms the SEC oversees. The 2024 amendments are now live for the whole covered market (larger firms December 3, 2025; all smaller firms June 3, 2026 — the requested extension was denied). They require written policies, an incident-response program, ongoing oversight of service providers, a 72-hour provider-to-firm breach-notice chain, and 30-day customer notice. Reg S-P is a named SEC exam priority for FY2026.

The annual, public list of what the SEC's exam staff will focus on in the coming year. It matters because it tells you, in advance, what an exam is likely to probe — and for FY2026, Reg S-P compliance is on the list, alongside scrutiny of firms' AI use and AI-washing claims.

The ongoing due diligence and monitoring of your vendors that Reg S-P now requires firms to maintain in writing — including building your incident-response program so a provider notifies you within 72 hours of a breach. It's the specific obligation an AI vendor file is built to satisfy.

AI tools your employees adopt and use without IT or compliance approval — a personal ChatGPT account, an unsanctioned meeting transcriber, a browser plugin. NobleCloak's view is that shadow AI is fundamentally an access problem, not a traffic problem: the risk is the data those tools were granted, which is why the answer is an OAuth inventory, not a firewall.

An AICPA examination report on a service organization's controls (security, availability, confidentiality, and more). Type I attests to control design at a point in time; Type II attests to operating effectiveness over a period. "No SOC 2 on file" is a real finding about a vendor. NobleCloak does not hold a SOC 2 report and our audit program has not started — we publish that, and every other claim about our own posture, at trust.noblecloak.com.

A Federal Reserve supervisory letter (April 17, 2026) that supersedes SR 11-7 on model risk. Notably, it carves generative and agentic AI out of model-risk scope (footnote 3) and is aimed at banks over $30 billion in assets, generally excluding smaller institutions. Read it as evidence of the regulatory vacuum, not as a mandate — a carve-out is not a rule.

The risk your organization inherits from the vendors and service providers it relies on — their security, their data practices, their failures becoming yours. AI has expanded this surface quietly, because vendors keep adding AI features to products you already bought.

The discipline of identifying, assessing, and monitoring the risk vendors pose — the program most regulated institutions already run and budget for. An AI vendor-risk assessment fits inside TPRM; it's a new question in an old process, not a new process. In the incumbent's 2026 survey, AI risk tied cybersecurity as institutions' top third-party concern for the first time.

The work of investigating a vendor's practices before and during an engagement — security, data handling, financial health, compliance posture. Under 07-CU-13 it is explicitly proportional: more scrutiny for critical, complex vendors, less for peripheral ones.