What FINRA SR 11-7 means for your LLM deployments
FINRA's 2026 Oversight Report explicitly flagged large language models as a model risk governance priority. SR 11-7, the federal guidance on model risk management, applies to algorithmic systems — and regulators have made clear that LLMs are algorithmic systems. Here's what that means in practice.
What is SR 11-7?
SR 11-7 is a 2011 guidance letter from the Federal Reserve and OCC titled "Supervisory Guidance on Model Risk Management." It defines a model as "a quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates."
Large language models fit squarely inside this definition. Regulators confirmed this in the 2026 Oversight Report and subsequent examination guidance. If your firm uses an LLM in any capacity that influences financial decisions, customer communications, or internal workflows, SR 11-7 applies.
The three pillars of SR 11-7 compliance are: (1) Model Development, Implementation, and Use— the model does what it's supposed to do; (2) Model Validation — an independent party confirms this; (3) Model Governance — policies, procedures, and documentation are in place and auditable.
Why LLMs make SR 11-7 harder
Non-determinism
Traditional quantitative models produce the same output for the same input. LLMs do not. SR 11-7 requires documenting model behavior — how do you document a stochastic system? LLMSec probes test for behavioral consistency across multiple runs.
Opaque failure modes
A credit scoring model fails in predictable ways. An LLM can be manipulated through natural language to deviate from its intended behavior in ways that are invisible to standard QA processes.
Vendor model risk
SR 11-7 Section 8 requires that you manage the risk of third-party models you rely on. Using GPT-4o-mini in production means GPT-4o-mini is subject to your model risk governance framework. "OpenAI does internal red-teaming" is not a valid model validation.
Validation independence
SR 11-7 requires that model validation be independent of model development. Running your own internal red team on a model you built does not satisfy this. An independent scanner with documented methodology does.
How LLMSec maps to SR 11-7 controls
Every LLMSec finding in a report includes the SR 11-7 section it addresses, the specific vulnerability class, and recommended remediation mapped to the control objective. Below is the mapping of probe categories to SR 11-7 sections:
| SR 11-7 Section | LLMSec Probe Categories | Risk Addressed |
|---|---|---|
| Conceptual Soundness | Jailbreak / role-play bypassConstraint overrideGoal hijacking | Model deviates from intended behavior under adversarial input |
| Ongoing Monitoring | Prompt injectionIndirect injectionSystem prompt override | Model behavior can be manipulated by untrusted user input |
| Outcomes Analysis | Toxicity / hate speechDiscriminatory outputBias detection | Model produces harmful or non-compliant outputs |
| Sensitivity Analysis / Stress Testing | Base64-encoded instructionsMulti-turn escalationAdversarial suffix injection | Model fails under adversarial edge-case inputs |
| Vendor and Third-Party | Data extraction / PII leakageSystem prompt leakageTraining data extraction | Model exposes confidential system data or fine-tuning content |
| Model Risk Governance | Reliability / refusal consistencyMulti-turn coherence | Model behavior is non-deterministic or inconsistent across sessions |
What an LLMSec report gives examiners
When a FINRA examiner asks "how do you validate your LLM models?", you need documentation. Here's what an LLMSec report contains that satisfies a model validation inquiry under SR 11-7:
The examination timeline you need to know
SR 11-7 issued
Federal Reserve and OCC establish model risk management framework for quantitative models at financial institutions.
Regulators clarify LLM scope
FINRA, SEC, and OCC issue guidance clarifying that SR 11-7 applies to ML and AI systems, including LLMs, used in financial contexts.
FINRA 2026 Oversight Report
LLM governance explicitly named as an examination priority. Firms without documentation of how they validate and govern LLMs are at examination risk.
Active examinations
FINRA examinations for 2026–2027 are ongoing. Examiners are asking about AI governance. Firms without a model validation process for their LLMs have material examination risk.
How is accuracy measured?Every LLMSec judge is calibrated against AllenAI's WildGuardMix benchmark. Measured precision, recall, F1, and false-negative rates are published — including weaknesses — on the calibration & methodology page. SR 11-7 Section 5 requires documented model validation methodology; we publish ours in full.