The Second Wave: AI's Sensitive-Data Exposure

Cyberhaven Labs' 2026 AI Adoption & Risk Report, built on billions of real data movements from 222 companies, paints a picture that has been underestimated: AI's "Second Wave" has arrived and is operational. 39.7% of all AI interactions involve sensitive data — the average employee inputs proprietary information into AI once every three days. Data is flowing to AI at an unprecedented scale.

The Second Wave: From Experimentation to Operational Dependency

The report exposes a sharp "AI Adoption Gap." Frontier organizations now use over 300 GenAI tools at nearly 6x the average adoption rate; 71.4% of employees use GenAI in leading organizations versus just 2.5% in cautious ones. The most aggressive adopters are technology (40.5%), pharmaceuticals (33%), and financial services (28.7%) — precisely the industries with the highest sensitive-data density.

When the adoption gap is this wide, data governance and security controls almost cannot keep pace. That is what distinguishes the "Second Wave" from early experimentation: it is no longer a few teams testing the waters, but an entire organization's data flows beginning to move through AI tools.

Sensitive Data: A Leak Risk Every Three Days

39.7% of all AI interactions involve sensitive data, meaning the average employee inputs proprietary information into AI once every three days. This is not an edge case — it is part of everyday workflow. With coding assistants covering nearly 50% of developers (90% in frontier companies) and 23% of enterprises having adopted agent-building platforms, both the channels and the frequency of sensitive data flowing into AI are climbing.

The risk is not only about input; it is about output and redistribution. Once proprietary data enters an AI tool, enterprises often lose control of it — who accessed it, what it was used for, and whether a third-party model trains on it may all fall outside the enterprise's control.

Shadow Accounts: One in Three Use Personal Accounts

One of the report's most troubling findings is that roughly one in three employees access AI tools via personal accounts, including 58% of Claude users and 60% of Perplexity users. This means a large volume of sensitive data flows to third-party systems outside enterprise control — data the enterprise cannot audit and cannot respond to quickly in the event of a breach.

Equally concerning: 82%of the top-100 most-used GenAI SaaS applications are classified as "medium" to "critical" risk. High adoption does not mean low risk — on the contrary, the most popular tools are often the ones that most need controls.

Chinese Open-Weight Models: Half of Endpoint Usage

Chinese open-weight models like DeepSeek and Qwen now account for 50% of all endpoint-based AI usage. For enterprises, this is a phenomenon that demands dedicated handling: substantial data is flowing to open-weight models whose security commitments, data residency, and auditability need separate evaluation — open source must not be assumed safe by default.

Combined with the shadow-account problem, enterprises face a double loss of control: personal accounts bypass enterprise governance, while open-weight models add data-governance uncertainty. Stacked together, the enterprise's sensitive-data exposure surface is significantly amplified.

An Action Checklist: Embed Data Governance into AI Workflows

Facing the "Second Wave," enterprises need to act on three fronts. First, build visibility — identify which of all AI interactions touch sensitive data; you cannot govern what you cannot see. Second, shut down shadow accounts — force AI use from personal accounts into enterprise-controlled workspaces with model toggles to prevent surprise activations. Third, govern high-adoption tools — apply dedicated DLP and least-privilege policies to coding assistants and agent platforms, ensuring model providers do not retain or train on enterprise data.

The key principle: data governance must be embedded into AI workflows, not retrofitted afterward. When nearly 40% of AI interactions already touch sensitive data, the cost of waiting for problems to surface is far higher than building controls at the point of adoption.

References:

• Cyberhaven Labs: 2026 AI Adoption & Risk Report
• Cyberhaven: 2026 AI Adoption & Risk Report (PDF)