Attackers exfiltrate through anything the agent can reach: allowlisted commands that make DNS lookups, web fetch tools, rendered links, and MCP servers with write access to public places. Each channel needs its own control, because a network proxy alone does not see tool calls. Test your setup with known exfiltration techniques before calling it safe.
Read in this order
- ArticleThe lethal trifecta for AI agents (Simon Willison)Start here: why any outbound channel plus private data is enough.
- ArticleGitHub MCP exploited: accessing private repositories via MCP (Invariant Labs)A real attack that leaked private repos through an MCP server.
- ArticleClaude Code: Data exfiltration with DNS, CVE-2025-55284 (Embrace The Red)Shows a quiet DNS channel you can replay against your own setup.