Shadow AI and the GLBA Safeguards Rule: Is Unsanctioned AI Use an Information Security Program Gap?
Plausibly, yes. The Safeguards Rule defines an information security program by reference to any handling of customer information. A prompt, upload or coding-assistant request that carries customer information to an unsanctioned AI tool is therefore a foreseeable risk that the risk assessment should identify and the safeguards should control. The rule does not mention AI, so this is an interpretation of the text, not a stated regulator position. A written policy alone will not show what AI was allowed to do or what it did.
Mapping shadow AI to program requirements
Why unsanctioned AI use falls within the rule
As of November 2026, 16 CFR Part 314 is the current FTC Safeguards Rule. It implements GLBA sections 501 and 505(b)(2) and sets standards for reasonable administrative, technical and physical safeguards that protect the security, confidentiality and integrity of customer information. It applies to financial institutions over which the FTC has rulemaking authority. Bank holding companies fall under a separate regime, the Interagency Guidelines in 12 CFR Part 225, Appendix F, which also require a written information security program. Confirm which regime governs your institution before mapping anything.
The scope question turns on the definition. The rule describes an information security program as the safeguards used to access, collect, distribute, process, protect, store, use, transmit, dispose of, or otherwise handle customer information. An employee who pastes a customer record into a public chat assistant, a browser extension that reads page content, or a coding assistant that receives a data sample is distributing or transmitting customer information to a third party. Whether the tool was sanctioned does not change that.
The rule is also method-neutral. The FTC's 2021 amendments require programs to address certain aspects of security but do not prescribe a particular technology. No tool is mandated for AI. Any control you choose has to be justified by your risk assessment and judged by whether it works.
Documented controls versus enforced controls
Three common responses each leave a gap when used alone. Policy documents and acceptable use statements describe intent. They do not show what happened in a specific interaction. Discovery tells you which AI tools are in use, which is necessary for the risk assessment, but it does not stop customer information from reaching them. After-the-fact review can detect and investigate, but only after the information has left.
The FTC's 2020 matter on vendor oversight shows the pattern examiners look for. The FTC alleged that a company wrote a third-party vendor risk management policy but did not follow through to ensure it was implemented. The complaint, which ended in a proposed settlement, alleged failures to oversee service providers, to identify reasonably foreseeable risks and to assess the sufficiency of safeguards. That matter predates the 12-month window and concerned vendors generally, not AI. The lesson still carries over: a written policy is weak evidence when implementation cannot be shown.
The practical distinction is the evidence each control produces. A policy produces a document. Discovery produces an inventory. Enforcement in the execution path produces a decision record for each interaction: which policy applied, what was allowed, redacted or blocked, and when. That record is what lets a CISO answer an examiner's question about a specific event.
Closing the gap in a defensible order
The sequence below keeps the response tied to the rule's risk-based structure instead of a tool purchase.
- Add AI to the written risk assessment: Include public assistants, AI-enabled sites, extensions and coding assistants as a risk category, and assess whether current safeguards are sufficient.
- Discover actual use: Establish which tools employees use and which receive customer information, so the assessment rests on observed behavior.
- Classify AI providers: Treat third-party AI services that receive customer information as service providers, with selection criteria and contract safeguards.
- Define interaction outcomes: Decide which interactions involving customer information are allowed, redacted, blocked or escalated for human approval.
- Enforce and retain records: Apply those decisions at runtime and keep a record of each one for examiner review.
- Test and reassess: Verify that written AI policy is implemented, and revisit the assessment when AI adoption materially changes operations or business arrangements.
Where runtime enforcement fits, and where it does not
Enforcement in the execution path sits between users or agents and the models, tools, MCP servers, AI sites and coding assistants they reach. Because it acts before customer information leaves, it can prevent what after-the-fact review can only detect. Its scope is the individual interaction or tool call.
It complements the other program elements and does not replace them. It does not perform the risk assessment, select or contract with service providers, run the testing program or manage incident response. It supplies safeguards that the risk assessment can point to and records that the testing and incident processes can draw on. Calibrate its use to the sensitivity of the customer information and the scale of your AI use, and document that reasoning in the assessment.
Ownership matters as much as tooling. Assign and document who owns AI risk, who approves exceptions, and who reviews escalations. Without that, enforcement produces decisions that no one is accountable for.
How Trussed supports this control layer
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Each interaction or tool call can be allowed, redacted, blocked, rate limited or escalated for human approval. Enforcement is at runtime, not reviewed after the fact. Enforcement generates continuous audit-ready evidence of what was allowed, redacted or blocked. Trussed can be deployed self-managed in the customer cloud, with configurable retention and data residency.
Trussed has run in continuous production in a regulated enterprise for more than 18 months and serves customers in insurance, pharma, healthcare and education. It addresses the enforcement and evidence layer described above. The risk assessment, service provider oversight, testing and incident response remain the institution's own program responsibilities.
From unsanctioned AI use to defensible evidence
Identify
Name AI tools, sites and assistants as a risk to customer information in the written risk assessment.
Map
Tie each exposure to the program element that should control it.
Enforce
Apply policy to interactions in the execution path: allow, redact, block, rate limit, escalate.
Evidence
Keep a durable record of each decision for examiners and auditors.
Review your AI exposure against the Safeguards Rule
Walk through how runtime enforcement and audit-ready evidence fit alongside your existing information security program.
Talk to an Expert