Report by Harmonic Security
GenAI Data Exposure: What GenAI Usage Is Really Costing Enterprises
Key Findings
15% of Google Gemini use by employees was via personal accounts.
26.3% of ChatGPT use by employees was via personal accounts.
13.7% of all sensitive prompts analysed in Q2 originated in Microsoft Copilot.
72.6% of all sensitive prompts analysed in Q2 originated in ChatGPT.
1.8% of all sensitive prompts analysed in Q2 originated in Perplexity.
Of these incidents involving Chinese GenAI tools, the exposed data types included: 32.8% involving source code, access credentials, or proprietary algorithms; 18.2% including M&A documents and investment models; 17.8% exposing PII such as customer or employee records; and 14.4% containing internal financial data.
Of analyzed prompts and files submitted to 300 GenAI tools and AI-enabled SaaS applications between April and June, 22% of files (totaling 4,400 files) and 4.37% of prompts (totaling 43,700 prompts) were found to contain sensitive information.
The average enterprise uploaded 1.32GB of files (half of which were PDFs) to GenAI tools and AI-enabled SaaS applications in Q2. A full 21.86% of these files contained sensitive data.
Code leakage was the most common type of sensitive data sent to GenAI tools.
7.95% of employees in the average enterprise used a Chinese GenAI tool.
535 separate incidents of sensitive exposure were recorded involving Chinese GenAI tools.
Sensitive data in files sent to GenAI tools showed a disproportionate concentration of sensitive and strategic content compared to prompt data, with files being the source of 79.7% of all stored credit card exposures, 75.3% of customer profile leaks, 68.8% of employee PII incidents, and ◦ 52.6% of total exposure volume in financial projections.
47.42% of sensitive employee uploads to Perplexity were from users with standard (non-enterprise) accounts.
In Q2, the average enterprise saw 23 previously unknown GenAI tools newly used by their employees.
5.0% of all sensitive prompts analysed in Q2 originated in Google Gemini.
2.5% of all sensitive prompts analysed in Q2 originated in Claude.
2.1% of all sensitive prompts analysed in Q2 originated in Poe.