diff --git a/docs/posts/2026-10-28-the-allowlist-that-lets-the-env-out.md b/docs/posts/2026-10-28-the-allowlist-that-lets-the-env-out.md index 08b8ee4..189833b 100644 --- a/docs/posts/2026-10-28-the-allowlist-that-lets-the-env-out.md +++ b/docs/posts/2026-10-28-the-allowlist-that-lets-the-env-out.md @@ -1,25 +1,3 @@ -## Submission drafts (remove before publishing) - -**r/netsec.** *Agent tool allowlists that still leak .env, and agents that -rewrite their own policy.* I maintain an open-source, fail-closed gateway -between coding agents (Claude Code and Codex hooks, MCP servers) and their -tools, and after presenting it I found two classes of bugs in my own code: an -allowlisted binary, origin or path that becomes the exfiltration channel -(`curl -d @.env`, the key in an allowed URL's query string, `.env.local`, -`cat .ENV` on macOS), and an agent that can edit the files that load the -firewall. The write-up has minimal reproductions with before/after output, why -verb- and path-based allowlists cannot catch these, the fixes in 0.4.0, and -what is still not covered. - -**Hacker News.** *Your agent's allowlist still lets the .env out.* Allowlisting -curl, an origin or a directory decides which tools a coding agent may use, not -which data may leave through them. This walks through the leaks I found in my -own agent firewall (curl's `@file`, a secret in an allowed URL, copies under -other names, a hook that forgot the session) and the agent rewriting the -settings that load the firewall, with the fixes and their limits. - ---- - # The allowlist that still lets the `.env` out *Valentín Torassa, 2026-10-28. Code: [VT-Agent-Firewall](https://github.com/ValentinTorassa/VT-Agent-Firewall), Apache-2.0.*