AI Compliance Framework
for SEC-Registered Advisers

Nine Controls

The framework

1. Governance, Risk, and Compliance

Establish an AI Risk Management Policy aligned with SEC cybersecurity rules and the NIST AI Risk Management Framework. Formally document approved AI use cases — client communication, internal operations, surveillance — and maintain an AI Bill of Materials (AIBOM) identifying datasets, open-source components, model versions, and training sources. Define governance roles across IT, compliance, and business units. Map AI usage against compliance risk factors including privacy, fiduciary duty, and supervisory obligations.

The AIBOM is the AI equivalent of a software bill of materials — a living document listing training inputs, model versions, and dependencies that is essential for understanding AI-related risk exposures during an incident or regulatory inquiry.

2. Access Controls

Enforce least privilege on AI models, APIs, and vector databases. Require multi-factor authentication for developers, administrators, and AI platform access. Implement audit logging of model usage, admin actions, and system prompts. Segment access between client-facing, internal-only, and test environments. No user — human or system — should have access beyond what their role strictly requires.

3. Data Protection

Never train AI on non-anonymized client data or NPI without an explicit, documented risk mitigation plan. Use differential privacy or data redaction where client data is fed into prompts. Secure vector databases and augmentation data with encryption in transit and at rest, change logging, anomaly alerts, and read/write segregation. Any third-party-hosted model should be reviewed under vendor due diligence, with legal clauses addressing training reuse, jurisdiction, data sharing, and supportability.

4. Deployment Strategy

The platform you choose determines what protections you have. The same underlying model offers very different data handling depending on which product you bought. Prioritize on-premise or isolated cloud environments for sensitive workflows. Block consumer-tier ChatGPT, Gemini, and Claude for any work-related use — these cannot satisfy Reg S-P service-provider oversight requirements. For any third-party-hosted AI, document contractual clauses on data usage, retention, training reuse, model tuning rights, and jurisdictional control.

5. Inference Security

Inference is where most users interact with AI — and where most real-world attacks occur. Prompt injections, jailbreak attempts, and adversarial inputs are active tactics used to override system safeguards. Implement input validation and prompt sanitization across all AI interfaces. Restrict models from responding to certain prompt types, implement contextual constraints, and filter outputs for hallucinations or policy violations before they are shared or stored. Monitor anomalous behavior — a model producing investment recommendations or revealing internal code logic should trigger an alert.

6. Monitoring and Auditing

Integrate AI usage logs into existing SIEM infrastructure so AI-related anomalies are visible alongside other system alerts. Track prompt refusals, failed inference attempts, and API usage spikes that deviate from baseline. Schedule red team exercises to simulate prompt injection and data leakage. AI systems should undergo periodic penetration testing, particularly those deployed in production or tied to sensitive workflows. Findings feed directly into AI risk assessments, policy updates, and training programs.

7. Staff Awareness and Training

The most sophisticated AI controls can be undone by human error. Incorporate AI-specific modules into annual cybersecurity training covering prompt injection, hallucinations, data leakage, and AI-generated phishing. Run tabletop exercises built around realistic AI incidents — receiving a convincing phishing email written by an LLM, or responding to a prompt that produces unauthorized investment advice. Provide clear policy-driven guidance on which AI use cases are acceptable and which are not. FieldCraft, available to BrainTrust members, includes modules designed specifically for AI use in regulated environments.

8. Model Lifecycle Management

AI is not a one-time install — it is a living system. Maintain model registries tracking lineage, dependencies, training sources, fine-tuning events, and deployment environments. Limit who can modify or deploy models, log every operation, and maintain rollback capability. Trained models that are integral to investment processes or client analytics should be treated as business-critical IP — their distribution should be restricted, monitored, and legally protected.

9. Regulatory Alignment

Include AI-specific threat scenarios in the annual cybersecurity risk assessment. Ensure the Cybersecurity Policies and Procedures Manual has a section dedicated to AI — describing where AI is used, how it is governed, and what controls are in place. Maintain a vendor register of all third-party AI tools, including open-source components, with due diligence documentation covering licensing, data retention, model provenance, and update mechanisms. Review AI systems as part of mock audits and SEC preparedness exercises.