Agents with shell or API access can leak data without rendering anything. They can encode secrets into a DNS lookup or post them through a legitimate tool such as a pull request or chat message. Allowlisted commands and trusted connectors are frequent blind spots.
Read in this order
- OfficialMITRE ATLAS AML.T0086: Exfiltration via AI Agent Tool InvocationStart here: the technique definition for exfiltration through agent tool calls.
- ArticleClaude Code: Data Exfiltration with DNS (CVE-2025-55284)Leaking data through DNS from an allowlisted command, with a CVE.
- ArticleInvariant Labs: GitHub MCP Exploited: Accessing private repositories via MCPA trusted GitHub connector used to leak private repositories.