Shadow AI’s Surge Threatens Data Privacy and Triggers New Regulatory Risks

Shadow AI is no longer a fringe concern

Verizon’s latest Data Breach Investigations Report (DBIR) reveals that 45 % of professionals now use generative‑AI tools at work, and 67 % of those users rely on personal, unsanctioned accounts. Compared with the previous year, the share of insider actions involving “shadow AI” – the practice of feeding corporate data into private AI services – has risen four‑fold across more than 22 000 breach incidents worldwide.

Employees are uploading source code, design documents, images, spreadsheets and even full research papers to platforms such as ChatGPT, Gemini, Claude, Grok and a host of niche coding assistants. In 28 % of data‑loss‑prevention (DLP) violations, source code was the material most often submitted, while 3.2 % of cases involved proprietary research documents.

Why regulators care

Under the EU General Data Protection Regulation (GDPR), personal data – which includes employee identifiers, client details and any information that can be linked to an individual – must be processed lawfully, transparently and with appropriate safeguards. When staff upload such data to an external AI model without a legal basis, the organization may be deemed a joint controller. Joint controllers share liability for any breach, and GDPR can impose fines of up to €20 million or 4 % of global annual turnover, whichever is higher.

The California Consumer Privacy Act (CCPA) and its amendment, the California Privacy Rights Act (CPRA), impose similar duties on businesses that collect or disclose personal information of California residents. Unauthorized transmission of that data to a third‑party AI service counts as a “sale” under the statutes, triggering the right of consumers to opt‑out and the possibility of statutory damages of $2 500–$7 500 per violation.

Other jurisdictions – such as Brazil’s LGPD, Canada’s PIPEDA and India’s PDPB – contain comparable provisions that can be invoked when shadow AI leads to cross‑border data flows without adequate contractual safeguards.

Immediate impact on users and companies

1. Loss of intellectual property – Uploading source code or design schematics to a public model can allow the provider to retain the data for future training, effectively handing competitors a free copy of proprietary assets.
2. Privacy breaches – Personal identifiers embedded in documents (e.g., employee IDs, client emails) become exposed to AI providers that may not be bound by the same privacy commitments.
3. Regulatory exposure – Failure to document the legal basis for processing, conduct a Data Protection Impact Assessment (DPIA) or secure appropriate data‑processing agreements can trigger enforcement actions.
4. Reputational damage – Public disclosures of shadow‑AI incidents erode trust among customers, investors and talent pools.

What organizations must change now

1. Enforce an AI‑specific Acceptable Use Policy

• Prohibit the use of personal accounts for any work‑related AI interaction.
• Require employees to obtain explicit approval before uploading any corporate data to an external model.

2. Deploy AI‑BOMs (AI Bills of Materials)

• Follow the open‑source framework released by Cisco’s AI‑BOM project to capture model version, training data provenance, prompts and configuration changes.
• Use AI‑BOMs to reconstruct the state of a model at the time of an incident, enabling faster forensic analysis.

3. Integrate DLP with generative‑AI monitoring

• Extend existing DLP rules to flag uploads of source code, PDFs, spreadsheets and images to known AI endpoints.
• Deploy proxy solutions that inspect outbound traffic for API calls to services such as api.openai.com, api.anthropic.com, etc.

4. Conduct regular Data Protection Impact Assessments

• Evaluate the risk of each approved AI use case against GDPR Art. 35 criteria.
• Document lawful bases (e.g., legitimate interests, performance of a contract) and retain records for supervisory‑authority review.

5. Update vendor contracts and Model‑Processing Agreements

• Ensure AI service providers commit to data minimisation, no‑re‑training of uploaded content, and prompt deletion after use.
• Include clauses that obligate the provider to comply with GDPR‑required safeguards such as Standard Contractual Clauses for any cross‑border transfers.

6. Educate the workforce

• Run mandatory training that explains the privacy implications of feeding data to AI tools.
• Provide approved, enterprise‑grade AI platforms that give security teams visibility into model usage.

The broader security picture

Verizon also notes that classic software vulnerabilities remain the leading cause of breaches, with the remediation rate for critical flaws from CISA’s KEV catalog dropping to 26 % in 2025. The median time to fully resolve a vulnerability has stretched to 43 days, underscoring that organizations cannot rely on patching alone to protect against the new vector introduced by shadow AI.

Ransomware continues to affect nearly half of all reported incidents, though the proportion of victims paying the ransom fell to 31 %, and the median ransom demand slipped to $139,875.

Bottom line for digital‑rights advocates

Shadow AI represents a dual threat: it jeopardises both corporate secrets and the privacy rights of individuals whose data may be hidden inside those secrets. By treating AI tools as data processors, applying existing privacy frameworks, and adopting transparent AI‑BOMs, companies can curb the surge, avoid punitive fines and uphold the rights of the people whose information they handle.

For a deeper dive into AI‑BOM best practices, see the Cisco AI‑BOM documentation.