In April 2023, JPMorgan Chase restricted employee access to ChatGPT. Goldman Sachs followed. Then Citigroup, Bank of America, Deutsche Bank, and Wells Fargo. Within 90 days, virtually every major financial institution on Wall Street had either banned or severely restricted the use of external AI chatbots. The stated reason was consistent across firms: the risk that proprietary information – trading strategies, client data, deal terms, compliance-sensitive material – would leak into AI training pipelines.

These were not precautionary overreactions. They were rational responses to a concrete threat. When a quantitative analyst describes a proprietary mean-reversion strategy in a ChatGPT prompt to help optimize the code, that description – the alpha signal, the entry criteria, the risk parameters – has been transmitted to a third-party server. Under OpenAI’s data practices at the time, consumer-tier conversations were eligible for model training. The analyst just handed a piece of the firm’s edge to a system designed to learn from everything it receives.

The financial services industry sits at the intersection of three regulatory frameworks – SEC, FINRA, and state/federal banking regulators – none of which were designed with AI data flows in mind. The resulting compliance challenge is the most complex in any regulated industry.

The Alpha Leakage Problem

In quantitative finance, alpha – the excess return above a benchmark – is the product. Hedge funds, proprietary trading firms, and asset managers spend hundreds of millions of dollars per year on research, data acquisition, and technology to develop strategies that generate alpha. The useful life of a trading signal depends on its exclusivity. The moment a strategy becomes widely known, it is arbitraged away.

This creates a unique risk profile for AI tool use. Unlike healthcare (where the risk is regulatory penalty) or legal (where the risk is privilege waiver), the risk in finance is direct economic destruction. A trading strategy disclosed through an AI system’s training pipeline does not just create a compliance problem – it destroys the asset.

The risk is not theoretical. Research published by Cornell University in 2024 demonstrated that large language models can extract and reconstruct structured information from training data with remarkable fidelity. The model memorization problem is well-documented: LLMs memorize and can regurgitate specific sequences from their training data. If a quant’s detailed strategy description enters a training pipeline, the probability that fragments of that description could surface in other users’ outputs is non-zero.

The financial industry’s response has been a barbell: either complete prohibition of external AI tools, or the development of internal AI systems running on proprietary infrastructure. The middle ground – using external AI tools with appropriate safeguards – is considered insufficiently secure for the most sensitive applications.

SEC Requirements and the AI Blind Spot

The Securities and Exchange Commission regulates the securities industry under a framework built around disclosure, record-keeping, and supervisory obligations. Several SEC requirements intersect directly with AI tool use.

Books and Records (Rule 17a-4)

SEC Rule 17a-4 requires broker-dealers to retain certain business records for specified periods. The rule covers communications “relating to the member’s business as such.” In 2023, the SEC began a massive enforcement campaign against firms for failing to preserve off-channel communications (primarily text messages and WhatsApp). Fines exceeded $2 billion across more than 40 firms.

AI interactions present the same risk. If a portfolio manager discusses investment strategy in an AI chatbot, that communication arguably relates to the firm’s business and is subject to retention requirements. Most firms have no mechanism to capture, archive, and supervise AI tool interactions. The record-keeping failure is immediate and systemic.

Regulation S-P (Privacy of Consumer Financial Information)

Regulation S-P requires financial institutions to protect the privacy of customer personal financial information. An analyst who enters a client’s portfolio details, account information, or financial situation into an AI tool is potentially violating Reg S-P’s safeguarding requirements. The regulation requires firms to adopt policies and procedures reasonably designed to protect customer records and information – and most firms’ policies were written before AI chatbots existed.

Proposed AI-Specific Rules

The SEC proposed Rule 15l-2 (for broker-dealers) and 211(h)-2 (for investment advisers) in July 2023, targeting the use of predictive data analytics and AI in interactions with investors. The proposals would require firms to eliminate or neutralize conflicts of interest arising from the use of AI technologies. While the final rules remain pending as of early 2026, the proposal signaled the SEC’s intent to regulate AI use directly, not just through existing frameworks.

The SEC has also indicated interest in AI model explainability – the ability to understand why an AI system made a specific recommendation. For firms using AI in portfolio management or trading, the black-box nature of modern LLMs creates tension with the SEC’s supervisory expectations.

FINRA’s Emerging Guidance

The Financial Industry Regulatory Authority, the self-regulatory organization overseeing broker-dealers, has been more proactive than the SEC in addressing AI-specific risks.

FINRA’s 2024 report on AI in the securities industry identified five primary risk areas: data privacy, model governance, bias and fairness, cybersecurity, and supervisory obligations. On privacy specifically, FINRA emphasized that firms must ensure AI tool use does not create unauthorized data sharing with third parties, and that firms’ supervisory systems must be capable of monitoring AI-assisted communications and decisions.

FINRA Regulatory Notice 24-09 (2024) specifically addressed the use of generative AI in communications with customers, requiring firms to supervise AI-generated content with the same rigor as human-drafted communications. A marketing email drafted by AI must be reviewed and approved under the same Rule 2210 framework as one written by a human. A research report summarized by AI must meet the same standards as one summarized by an analyst.

The practical challenge is supervision at scale. If an analyst uses AI to draft 50 research summaries per day, the supervisory burden quintuples. FINRA’s guidance implicitly assumes that firms will develop AI-aware compliance tools – using AI to supervise AI, creating a recursive dependency that raises its own questions about reliability and audit trails.

Bloomberg Terminal AI: The Infrastructure Question

The integration of AI into Bloomberg Terminal – the financial industry’s dominant information platform, with over 350,000 subscribers globally – illustrates the infrastructure-level stakes.

Bloomberg has deployed its proprietary BloombergGPT and subsequent models trained specifically on financial data. The key architectural decision: Bloomberg’s AI features process data within Bloomberg’s existing infrastructure, subject to its existing data handling agreements with clients. Data entered into Bloomberg’s AI features does not leave Bloomberg’s environment for third-party processing.

This is the model that most financial institutions prefer – a vertically integrated AI capability where data remains within a trusted perimeter. But it creates a competitive moat that reinforces Bloomberg’s market dominance. Smaller financial data providers and fintech startups cannot replicate this architecture at scale, and their use of third-party AI APIs introduces the data leakage risks that Bloomberg’s model avoids.

The Bloomberg approach also highlights a fundamental tension: the most capable AI models (GPT-4, Claude, Gemini) are developed by general-purpose AI companies with broad training data needs. Domain-specific models like BloombergGPT trade capability for data isolation. Financial firms must choose between the most capable model and the most secure architecture – a choice that Stealth Cloud infrastructure is designed to eliminate.

Proprietary Data in the AI Supply Chain

The AI data supply chain in finance extends beyond direct tool use. Financial firms face data leakage risk through:

Third-party vendors: Financial firms use hundreds of vendors for analytics, compliance, operations, and technology. When those vendors adopt AI tools – feeding client data into AI systems for processing – the financial firm’s data enters AI pipelines without the firm’s direct involvement. A 2025 survey by Deloitte found that 73% of financial services vendors planned to integrate generative AI into their products, but only 34% had updated their data processing agreements to address AI-specific data handling.

Employee behavior: Despite corporate restrictions, individual employees use consumer AI tools. The Samsung ChatGPT incident – where Samsung engineers pasted proprietary source code into ChatGPT – has direct parallels in finance. A compliance analyst drafting a suspicious activity report, a banker modeling a deal structure, a risk manager analyzing portfolio exposures – all of these tasks generate temptation to use AI tools that may not be sanctioned by the firm.

Alternative data providers: The alternative data industry – which supplies satellite imagery, web scraping, social media sentiment, and transaction data to hedge funds – is rapidly integrating AI into its data processing pipelines. Hedge funds that provide proprietary data to these providers for backtesting or validation may find that data incorporated into AI models that are then sold to competitors.

The concept of zero-persistence architecture addresses these risks at the infrastructure layer. When the AI processing system is architecturally incapable of retaining data beyond the immediate inference request, the supply chain leakage risk is neutralized at its root.

The Compliance Burden: What Financial Firms Need

For financial institutions operating under SEC, FINRA, and prudential banking regulations, compliant AI use requires:

Data Governance Framework

  • Classification of data by sensitivity tier (public, internal, confidential, restricted) with AI tool use permissions mapped to each tier
  • Inventory of all AI tools in use across the organization, including shadow AI detection
  • Data lineage tracking for any data that enters AI systems, enabling the firm to demonstrate where data went and what happened to it

Vendor Management

  • AI-specific amendments to vendor agreements, addressing data retention, training data use, model access, and data breach notification
  • Regular assessment of vendor AI practices – not just at onboarding, but continuously, as vendor practices evolve. The framework described in the AI due diligence checklist is directly applicable
  • Right-to-audit clauses that cover AI data handling specifically

Supervisory Controls

  • AI tool use monitoring systems integrated with existing compliance surveillance
  • Automated detection of sensitive data (account numbers, trade details, client names) in AI tool interactions
  • Escalation procedures for AI-related data incidents, integrated with existing SAR (Suspicious Activity Report) and regulatory notification processes

Record Retention

  • Capture and archival of AI tool interactions that constitute business communications under Rule 17a-4
  • Metadata preservation (timestamp, user, tool, data classification) even where conversation content is not retained
  • Integration with existing e-discovery and regulatory examination processes

The financial industry’s AI challenge is ultimately about the tension between competitive advantage and regulatory compliance. The firms that generate the most alpha from AI will be those that find architectures allowing them to use frontier models with frontier data – without that data leaking to competitors, regulators, or AI providers’ training pipelines. The AI provider privacy scoreboard provides a starting framework for evaluating which providers are architecturally compatible with this requirement.

The Stealth Cloud Perspective

In finance, data is alpha and alpha is revenue. Every prompt that reaches a training pipeline is a potential transfer of competitive advantage from the firm that generated the insight to the AI provider that absorbed it. The only architecture that preserves proprietary value is one where the AI system is structurally incapable of learning from the data it processes – where inference happens in isolation and nothing persists beyond the response.