Home / Glossary / AI data leakage
Glossary
What is AI data leakage?
AI data leakage is sensitive company data leaving the company through AI tools and AI agents. It happens when a prompt carries customer data, when an agent reads a secret, or when a tool call posts to an outside host. It is hard to see because it happens on the device and inside the agent, and it does not look like a file upload.
By the AIBound team · Updated 17 November 2026
Key takeaways
- AI data leakage is sensitive data leaving the company through a prompt, an AI agent or a tool call.
- An AI agent can read whatever its employee can read, including keys, source code and customer records.
- The most direct way to stop it is a rule that applies at the moment the agent acts.
Why AI data leakage matters
Employees use AI to work faster, and the fastest way to get help is to give the AI the real material. That means the failing code, the customer ticket or the contract draft. Once that data sits in an AI tool the company does not manage, the company no longer controls where it goes or how long it is kept.
AI agents raise the stakes. A coding agent on a laptop can open .env files, SSH keys, cloud tokens and customer records, because it runs with its employee's access. It can then post to a host, run a command or push to a repository. One leaked credential can open a much larger system than the file it came from.
How AI data leakage happens
Most AI data leakage starts with ordinary work. Nobody sets out to leak anything.
- 1 An employee or an agent picks up sensitive data, such as a pasted customer record or a file the agent reads on its own.
- 2 The data goes into a prompt, a tool result or a command.
- 3 The AI tool or agent sends it on, to a model provider, an outside host or another system.
- 4 The data now sits outside the company, and the company often has no record that it left.
AI data leakage vs traditional data leakage
Traditional data leakage is data leaving by email, file transfer, web upload or a lost device. AI data leakage is data leaving through a prompt or through an action an agent takes. The data is the same. The route is different, and so is the place where a control has to sit.
| Question | AI data leakage | Traditional data leakage |
|---|---|---|
| How does the data leave? | In a prompt, a tool call or a command an agent runs | In an email, a file transfer or a web upload |
| Who moves it? | An employee or an AI agent acting on its own | A person |
| Where can it be stopped? | On the device, inside the agent, as the action happens | On the network and in the channels the data crosses |
Sources
- OWASP LLM02: Sensitive Information Disclosure · genai.owasp.org
Sources checked 19 November 2026
Examples
AI data leakage examples
FAQ
Frequently asked questions
What is AI data leakage?
AI data leakage is sensitive data leaving the company through AI. It can be an agent that reads a secret, a prompt that carries customer data, or a tool call that posts to an outside host. It is hard to see because it happens on the device and inside the agent, and it does not look like a file upload.
What kinds of data leak through AI?
Six kinds leak most often. Source code and technical documents. Regulated personal data such as customer and health records. Roadmaps, contracts and other commercial material. Credentials, keys and secrets. Screenshots and scanned files. The sixth kind is any data an agent moves from one system to another without a person typing anything.
How do you prevent AI data leakage?
Start with the AI that can do the most damage, which is the agents that read files and run commands. Set rules that apply as the agent acts, such as denying reads of keys and secrets and denying posts to hosts you have not approved. Keep your current DLP for email, files and uploads.
Does AIBound read what employees type into AI tools?
No. AIBound enforces inside the agent on the device, and it does not work inside the browser page. In the browser it can detect AI use, but it does not read page content and does not block there. The AIBound Secure Client collects tool metadata only, with no PII or PHI.