What is data leakage?
Data leakage is the unintended exposure of sensitive or confidential data to a person, system or location that should not have it. Unlike a breach, it usually happens through normal work, such as a pasted prompt, a misconfigured sharing link or a log file, rather than through an attack.
For example, an engineer debugging a claims workflow copies a failing request into a team chat channel. The request holds a member’s name, date of birth and diagnosis code. Nobody stole anything, but protected health information now sits in a chat history that hundreds of people can search and that the chat vendor stores.
Common sources of data leakage
- Prompts and AI chat. Employees paste customer records, contracts or code into AI tools.
- Logs and traces. Applications write full requests, including personal data, to log files that many teams can read.
- Sharing links and storage settings. A folder or bucket set to public, or a link forwarded outside the company.
- AI agent tool calls. An AI agent sends real values to a third-party tool or another AI agent.
- Test and analytics copies. Production data copied into lower environments with weaker access rules.
- Email and file transfer. Attachments sent to the wrong recipient or a personal account.
Data leakage vs data breach vs data loss
| Term | What happens | Typical cause |
|---|---|---|
| Data leakage | Data is exposed to people or systems that should not have it | Everyday work, weak process, misconfiguration |
| Data breach | Confirmed unauthorized access to data, often with legal notice duties | An attacker or insider acting on purpose, or a leak that someone exploits |
| Data loss | Data is destroyed or unavailable | Deletion, hardware failure, ransomware |
Data loss prevention (DLP) tools use “loss” to mean data leaving the organization, which is closer to leakage than to destroyed data.
Why data leakage grows with AI systems
AI systems move data between systems on their own. A single prompt can include retrieved records from several sources. An AI agent can pass those values to tools, memory and other AI agents in seconds, and an answer can show values to a user who has no right to see them. Each step is a new place for data to leak. LLM data leakage covers the model-specific cases, and the guide on how to stop AI agents and MCP from leaking PII shows the fixes.
How to reduce data leakage
- Find where sensitive data lives and where it flows.
- Send only the data a task needs.
- Mask or tokenize sensitive values before they leave a trusted system or reach a log.
- Control who sees real values, by role and by context.
- Log access and review it.
How Protecto helps
Protecto reduces data leakage in AI workflows. It detects sensitive values in prompts, retrieved context, tool outputs and responses and replaces them with consistent tokens before they reach the model or a log. Users and tools without access see the data masked, while authorized users receive real values where policy allows. Protecto works alongside your existing DLP. See data leak prevention for AI.
Frequently asked questions
What is the difference between data leakage and a data breach?
Data leakage is unintended exposure, often through everyday work. A data breach is confirmed unauthorized access, usually by an attacker, and often triggers legal notification duties. A leak becomes a breach when someone unauthorized accesses the exposed data.
What is data leakage in machine learning?
In machine learning, data leakage also names a modeling error: information from the test set or from the future slips into training data, so the model looks more accurate than it is. It is a different problem from exposing sensitive data.
Can encryption prevent data leakage?
Only in part. Encryption protects stored and transmitted data, but leakage usually happens after data is decrypted for use, for example when it is pasted into a prompt or written to a log.
Related terms
Data loss prevention (DLP) · AI DLP · Sensitive data · Unauthorized data access, use or transfer · Logs