AI data security: Best practices for protecting enterprise data from AI-driven exposure

The most common risks fall into a few categories:

• Shadow AI: Employees using unsanctioned public AI tools with sensitive work data.
• Prompt and upload leakage: Pasting source code, PII, or contracts into chatbots that retain or train on inputs.
• Prompt injection: Attackers hiding malicious instructions inside documents, emails, or webpages so that AI assistants follow the attacker's commands instead of the user's.
• Over-permissioned AI assistants: Copilot-style tools inheriting access to high-security folders.
• Training data poisoning and tampering: Malicious or unvetted records corrupting model behavior.
• Compliance exposure: Data movement into AI tools that falls outside the scope of the GDPR, the EU AI Act, HIPAA, the PCI DSS, and similar AI frameworks.

Enterprises should align with voluntary frameworks for internal AI governance and comply with binding regulations that carry legal force.

Frameworks:

• NIST AI Risk Management Framework (AI RMF): Voluntary framework for governing, mapping, measuring, and managing AI risks.
• OWASP Top 10 for LLM Applications: Application-level risks like prompt injection, insecure output handling, and training data poisoning.
• MITRE ATLAS: Adversarial threat knowledge base for AI systems, modeled on MITRE ATT&CK®.
• CISA/NSA Joint Guidance on AI Data Security (2025): Government guidance specifically on securing data used to train and operate AI.
• EU AI Act: Binding regulation in the European Union (EU) that classifies AI systems by risk and imposing obligations on high-risk deployments.
• ISO/IEC 42001: Certifiable management system standard for AI, useful for organizations that need formal compliance attestation.

The EU AI Act classifies AI systems by risk level—unacceptable, high, limited, and minimal—and imposes obligations proportional to that risk. The most relevant requirements for enterprises apply to high-risk AI systems and include:

• Maintaining technical documentation of the AI system's design, intended purpose, and limitations.
• Ensuring data governance and quality for training, validation, and testing datasets.
• Logging system activity for traceability and post-incident analysis.
• Enabling human oversight throughout the AI life cycle.
• Registering the system in the EU's public AI database.

Providers and deployers of general-purpose AI models face additional transparency and copyright obligations. Non-compliance can attract fines of up to 7% of global annual turnover, comparable in scale to GDPR penalties.

ChatGPT Enterprise and Microsoft 365 Copilot are significantly safer than their consumer counterparts but require careful deployment. Neither products uses prompts to train foundation models, keep data inside the organization's tenant, and inherit existing identity and compliance controls. They do not, however, eliminate every risk. Microsoft's own Copilot deployment guidance recommends auditing SharePoint permissions and restricting the search to limit discoverability of sensitive content. Enforcing least-privilege access, auditing folder permissions before rollout, and monitoring usage can help reduce such unintended exposure.

Shadow AI is the use of AI tools, plug-ins, or AI features inside SaaS applications without IT approval or visibility. It includes employees using personal ChatGPT accounts on corporate devices, browser extensions that relay page content to AI services, and AI features silently activated inside already sanctioned SaaS tools.

To detect it:

• Monitor endpoint and browser activity for connections to known AI services.
• Inspect prompts and file uploads at the endpoint, not just at the network perimeter.
• Analyze web traffic and proxy logs for traffic to AI domains and APIs.
• Audit browser extensions and IDE plugins on managed devices.
• Establish behavioral baselines to surface anomalies—sudden spikes in AI tool use, uploads outside business hours, or access from unusual locations.