Skip to content

Security: SocialGouv/iterion

SECURITY.md

Security Policy

Supported versions

iterion ships from a single mainline. The latest released minor version receives security fixes; older minors are only patched when an operational impact is documented and reproducible.

Version Supported
Latest minor (see releases) Yes
Previous minor Critical fixes only, best-effort
Older No

Reporting a vulnerability

Please use GitHub's private Security Advisory workflow:

  1. Go to the repository's Security tab → Report a vulnerability.
  2. File a private advisory describing the impact, the affected version, and a minimal reproducer if you have one.

This routes the report to the maintainers without going through public issues. We aim to acknowledge within 3 business days and, for confirmed issues, ship a patch within 30 days for high-severity findings and 90 days for medium-severity.

We will credit reporters in the released advisory unless you ask otherwise.

Threat model & scope

iterion is a workflow engine that drives LLM backends (Anthropic, OpenAI, …) and executes shell tools the operator authorises. The in-scope surface for vulnerability reports:

  • Auth / identity (pkg/auth/, pkg/identity/, pkg/secrets/): authentication bypass, privilege escalation, tenant isolation failures, secret exfiltration.
  • Server HTTP / WS (pkg/server/): unauthenticated access to authenticated endpoints, cross-tenant data leaks, CSRF on state-changing routes, injection via untrusted headers/queries.
  • Sandbox (pkg/sandbox/): container escape, allowlist bypass on the network proxy, mount-traversal from inside the sandbox.
  • Vendor supply chain (vendor/): malicious dependency, broken pinning, untrusted code path.
  • Bundle handling (pkg/bundle/): tar/zip extraction issues, symlink-traversal, path-traversal.
  • Cloud deployment (charts/, pkg/cloud/, Dockerfile): privileged container, exposed admin endpoint, secret in values.yaml example.

The following are out of scope unless they amount to one of the above:

  • DSL workflows that authorise dangerous tools when the operator explicitly configures them — that is by design.
  • Theoretical denial-of-service requiring an authenticated operator account.
  • Misconfiguration in the operator's deployment (broad NetworkPolicy, weak passwords, leaked credentials) that does not stem from a defective default in this repo.
  • Issues only reachable when iterion is run on a custom fork.

Coordinated disclosure

We ask for a 90-day non-disclosure window starting from acknowledgement for high-severity findings, shortened by mutual agreement once a fix is released. Pre-disclosure to downstream operators (managed deployments, notable adopters) happens on a need-to-know basis the maintainers coordinate.

Hardening guidance

The docs/cloud-public-exposure-checklist.md captures the operator-facing hardening steps that complement this policy (TLS, NetworkPolicies, secret rotation, audit logging).

There aren't any published security advisories