This project is experimental. Releases are 0.x and make no stability promises:
fixes land on main and ship in the next release, and only the latest release and
the latest commit are supported.
Please report security issues privately through GitHub's "Report a vulnerability" form on this repository's Security tab. Do not open a public issue or pull request with details. If that option is not available to you, open an issue titled "Security contact request" with no details and a private channel will be arranged.
Include what you found, how to reproduce it, and what you think the impact is. This is a spare-time project: expect an acknowledgement within a few days and a best-effort fix. There is no bounty programme. Credit is given if you want it.
- Inbound mail authentication: accepting mail that should be rejected, or bypassing the sender allowlist, the DKIM and SPF binding, or the forwarded-mail (ARC) checks.
- Authorization: an agent reaching a lane it was not granted, a revoked agent or a revoked share link still working, or token handling flaws.
- Disclosure: anything that leaks message content, secrets, tokens, link URLs or deployment details through the Worker, its logs, error messages or the CLI.
- Resource exhaustion from a single message: zip bombs, parser blow-ups and similar inputs that defeat the documented limits.
- The repository's own build, CI and deploy scripts, including anything that could leak a credential or personal deployment values.
- Prompt injection through message content. Email bodies are untrusted by design and the project does not try to detect malicious instructions in them. What a consuming agent does with that text is the operator's responsibility (see the threat model). A way to make untrusted content appear in a field that is presented as trusted is in scope.
- Volumetric floods of the lane address and platform quota exhaustion. The mitigations are documented in docs/security.md.
- Vulnerabilities in Cloudflare, your mail provider, or MCP clients and agents.
- Misconfiguration the documentation warns against, such as listing a shared mail provider's domain as an allowed DKIM domain, or using an over-broad API token.
- Reports that need the secret lane address together with access to your Cloudflare account or secrets.
The code is only part of the picture. See docs/security.md for the Cloudflare rules worth putting in front of the Worker, and for the known limits of the design. Practical basics: give the CLI a narrow API token, create agent tokens yourself in a terminal, grant each agent only the lanes it needs, rotate a lane's address if it leaks, and keep consuming agents narrowly scoped.