AI glossary · Basic · AI security · Agents
Data exfiltration
Data exfiltration is when an unauthorized person or program secretly moves data from inside an organization to the outside. In the AI era, a new risk has emerged: AI agents being tricked into sending data out on their own.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Security & safety
- Fields
- AI security, Agents
- Added
- 2026-10-09
How it works
Traditional data exfiltration often involved hacking into a server and copying files. In AI services, attention has turned to a path through indirect prompt injection. An attacker plants hidden instructions in a web page, email or document, and an AI agent that reads it follows those instructions to find and read internal data. It then appends the data to a link URL or image URL, or calls a tool such as email sending, to push it out to the attacker's server. Because the user sees only an ordinary answer on screen, the leak is hard to notice.
Why it matters
AI agents often have access to several tools at once, such as email, files, databases and web search. When three conditions come together in one agent — 'can access sensitive data,' 'reads untrusted text' and 'can send data outside' — the risk of exfiltration rises sharply. If customer information or trade secrets leak, the result is not only financial loss but also legal liability and lost trust. That is why, as AI adoption grows, preventing data exfiltration is becoming a core task for security teams.
What to keep in mind
Hoping the model will refuse attack instructions is not enough. The basics are to reduce the data an agent can access according to the principle of least privilege, and to restrict outbound channels (external links, email sending, network requests) with an allowlist. Sensitive actions should be preceded by a human-in-the-loop step where a person confirms them, and logs that monitor outgoing data should be kept. Note also that the Korean term for 'data exfiltration' is sometimes used broadly to include accidental disclosure of information (a leak), so it is worth checking the meaning from context.
Example
Among the news featured on this site was a report that customer information at Yegaram Savings Bank and Welcome Savings Bank was leaked in an 'AI hacking' incident targeting financial firms (no additional damage to the other member companies had been confirmed, according to the report). In finance, customer information is itself a core asset, so data exfiltration is especially sensitive. When companies deploy internal AI assistants, they prepare by blocking automatic display of external links and designing tools such as email sending to run only after user confirmation.