The Data Defense Dilemma: Protecting Trade Secrets from Internal AI Leakage

The rapid adoption of generative AI tools—ranging from expansive Large Language Models (LLMs) to specialized code assistants—has introduced critical new vulnerabilities into modern corporate data security.  Today, the primary risk to intellectual property is no longer limited to malicious external hacking; it is increasingly driven by inadvertent internal disclosure.  This occurs when employees feed proprietary data into public-facing models, allowing sensitive information to be absorbed by an AI engine and regurgitated to outsiders, competitors, or malicious actors.  To maintain a competitive edge and satisfy evolving legal requirements, companies must establish clear and actionable policies to address this threat.  Failure to update security measures against modern AI usage jeopardizes legal protection under frameworks like the Defend Trade Secrets Act (DTSA), which requires companies to demonstrate that they have taken “reasonable measures” to safeguard their IP, including reasonable safety precautions for internal AI tools.

The New Frontier of Intellectual Property

In this era of rapid AI model iteration, trade secret protection is increasingly favored over traditional patents.  Trade secrets provide immediate, potentially perpetual protection for assets that are difficult to patent, evolve too rapidly for traditional filing, or lack a “human creator.”[1]  Patents also require full public disclosure—an undesirable trade-off in the fast-moving, difficult-to-detect field of AI.  Thus, trade secret protection is gaining importance.

A robust IP defense strategy must focus on two primary categories of proprietary assets: the Core AI Model and its Training Data.  

• Core AI Models: Encompasses specific architectures, unique parameters, and proprietary algorithms.
• Training Data: Often considered a company’s most valuable asset, encompassing cleaned, structured, proprietary datasets, specialized internal workflows, confidential prompts, and deployment strategies.  

Because AI technology evolves so rapidly, trade secret law has become a vital mechanism for preserving training data as a competitive advantage.

Understanding the Threat Landscape

Risks to trade secrets generally manifest across four main vectors: human error, use of “Shadow AI,” malicious actors, and output exposure.

Human Error: The most common point of failure remains the employee.  When proprietary information, such as source code, internal documents, or customer lists, is pasted into public-facing AI services, external providers may use that input data to train their foundational models.  Beyond direct copy-pasting, companies face risks from prompt injection (where attackers manipulate an LLM to override its guardrails to reveal internal data) and supply chain vulnerabilities (where third-party components contain exploits designed to exfiltrate information).

Shadow AI: Beyond enterprise tools, organizations face a growing risk from “Shadow AI”—the unsanctioned use of personal AI accounts, unapproved browser extensions, and consumer-grade tools on personal devices.  Because consumer services typically default to using user inputs for model training, Shadow AI bypasses standard enterprise logging and endpoint monitoring, quietly exposing trade secrets to third-party providers without administrative visibility.

Malicious Actors: On the technical front, attackers utilize sophisticated methods to extract the underlying architecture and training sets of proprietary models.

• Model Extraction: Repeatedly querying a deployed model to infer or reconstruct its proprietary weights and algorithms.
• Membership Inference Attacks (MIA): Determining if a specific data point was included in a training set.
• Model Inversion Attacks: Reconstructing sensitive training data—such as personally identifiable information (PII)—from AI output.

Output Data Exposure: IP leakage extends beyond input data; output data also poses severe legal risks.  If an employee incorporates AI-generated code or text trained on third-party trade secrets or copyrighted code into a core product, the company risks downstream IP infringement liability.  Further, because AI-generated output lacks copyrightability under US law due to the “human authorship” requirement, relying heavily on raw AI output risks introducing proprietary assets into the public domain.

The Three Pillars of Trade Secret Protection

To counter these diverse threats, a comprehensive defense must rest upon three pillars: Policy & Legal, Human Training, and Technical Controls.  To streamline implementation, do not treat “AI” as a monolith.  Segment and protect its underlying components individually, including training data, model weights, prompts, and outputs.

Policy and Legal: This pillar establishes a legal foundation demonstrating “reasonable measures” of protection:

• Asset Classification and Audits: Conduct annual trade secret audits to classify AI assets and conspicuously label sensitive documents with warnings (e.g., “Confidential/Trade Secret – Do Not Upload to External AI”).
• Updated Agreements: Revise Non-Disclosure Agreements (NDAs) and Invention Assignment Agreements (IAAs) to specifically prohibit employees from submitting company IP assets to external LLMs.  
• Vendor Contract Enforcement: Ensure enterprise vendor contracts include strict non-retention clauses and forbid training on company inputs.  Enterprise agreements should include explicit vendor representations and indemnification against third-party IP infringement claims, as well as explicit definitions of “data ownership” covering fine-tuned weights, embeddings, and customized prompt libraries.  These materials should be considered the exclusive, confidential property of the company.
• User Terms: Explicitly prohibit reverse engineering, data scraping, and “knowledge distillation” from output data in all user agreements to establish “improper means” for future litigation.
• Private Environments: Implement a “walled gardens” policy, which shifts sensitive development to “Private LLMs” or air-gapped environments where inputs are never used to train external, third-party models. 
• Patent Hedging: Meticulously document human interventions and creative decisions to preserve future patenting options.

Human Training & Culture: The second pillar builds a “human firewall.”

• Clear Usage Policies: A zero-tolerance policy must be enforced regarding the uploading of proprietary data to general-purpose AI services.
• Continuous Education: Mandatory training is essential to define what constitutes a trade secret, including unique workflows and financial projections.
• Offboarding Protocols: This protection continues even after an employee leaves; thus, specialized exit interviews must reinforce perpetual confidentiality obligations, and all system access and API keys must be revoked immediately.

Technical Controls: The third pillar deploys automated systems to enforce security policies.

• Access Control: Implement Role-Based Access Control (RBAC) and a “Zero-Trust” architecture for all AI assets.  
• Data Loss Prevention (DLP): DLP tools should be used to block the transfer of proprietary data to external API endpoints, paired with automated labeling of trade-secret materials.  Mandate the use of internal, isolated, or private LLM instances.
• Advanced Privacy Guardrails: For highly sensitive assets, deploy advanced techniques like Federated Learning and Differential Privacy to prevent attackers from reconstructing training data.
• Context-Aware RAG Governance: Companies using Retrieval-Augmented Generation (RAG) to ground LLMs in internal data create new internal leakage potential.  RAG systems convert corporate documents into embedded vectors stored in specialized databases.  These require context-aware access controls tied to the original document permissions; otherwise, an employee querying an internal RAG assistant could inadvertently retrieve confidential data or trade secrets that they lack permission to view.

Ultimately, protecting IP is an ongoing process that requires continuous security audits, model monitoring, and a dedicated incident response plan.  These three pillars allow businesses to confidently leverage the transformative power of AI while safeguarding the data that fuels their competitive edge.

Navigating Regulatory Conflicts

As companies move forward, they must navigate a delicate balance between legal security and compliance.  New regulations, such as the EU AI Act, frequently demand disclosures about training data or model safety, creating a direct conflict between regulatory transparency mandates and trade secrecy.  When disclosure is legally required, information must be conspicuously labeled and designated as a trade secret to maximize legal protection.  

Conclusion

While GenAI presents unprecedented challenges to human-centric law, it is not the death of trade secrecy.  Instead, it forces trade secret law back to its foundational purpose: protecting valuable proprietary information from devious or negligent exposure.  By pairing technical controls with clear legal strategies, businesses can confidently leverage generative technologies while maintaining the core assets that drive their competitive advantage.

[1] Under U.S. patent law, an inventor must be a natural person.  35 U.S.C. 100(f); Thaler v. Vidal, 43 F.4th 1207, 1211 (Fed. Cir. 2022); MPEP 2109 Inventorship [R-01.2024].