Data exfiltration through agents
Sensitive data leaves through the agent: in a tool call, a message, a link or an image URL the client loads. It usually follows an injection.
- LLM02:2026 Sensitive Information Disclosure
- ASI02:2026 Tool Misuse and Exploitation
- AML.T0086 Exfiltration via AI Agent Tool Invocation
- AML.T0077 LLM Response Rendering
- AML.T0136 Exfiltration via AI User Interface
How it works
- 1An agent that can read private data and also reach the outside world can be made to connect the two.
- 2The data can leave in plain sight, in a URL parameter, a rendered image, an email or a file the agent writes somewhere public.
- 3Because the action uses the agent's own permissions, access controls see an authorised request.
Real cases on the map
- CriticalMCP Path Traversal Allows Full-Filesystem Read and JWT Secret Exfiltration Leading to Account Takeoverlangflow, langflow-base
- Critical@zereight/mcp-gitlab: Unauthenticated arbitrary file read via `upload_markdown` enables PAT exfiltration and full account takeover@zereight/mcp-gitlab
- CriticalSSRF + LLM API key exfiltration via user-controlled basepath / baseUrl in 8 chat-model nodesflowise, flowise-components
- Criticalgemini-mcp-tool vulnerable to OS command injection and @file exfiltration via prompt quoting (CVE-2026-0755)gemini-mcp-tool
- CriticalMLflow: Environment variable injection in AI Gateway secrets enables server-side credential exfiltrationmlflow
- HighMCP Atlassian: Arbitrary File Read & Exfiltration (Confused Deputy) in JIRA update_issuemcp-atlassian
Published advisories whose own titles describe this technique.
What to watch for
- Outbound requests to new domains after private data was read
- Private data in URLs, image links or messages
- Unusual volume in tool outputs leaving the system
How to defend
- Separate reading private data from sending data out, or require approval to combine them.
- Restrict outbound destinations and rendering of remote content.
- Watch for data in URLs and links.