ContextFirewall sits on the path between an AI agent and its memory, and its job is to keep secrets and untrusted facts out of the model's context. We take security reports seriously.

The project is pre-1.0. Only the latest commit on main (and the hosted demo built from it) receives security fixes.

Please do not describe a security problem in a public issue, discussion, or pull request.

Report it privately instead:

1. Use GitHub's Report a vulnerability form (Security tab → Advisories).
2. If that form is unavailable, open a new issue titled "Security contact request" that contains no details about the problem. The maintainer, @himanshu748, will reply with a private channel for the report.

Please include:

• the affected component (backend/, frontend/, mcp/, supabase/) and commit SHA,
• steps to reproduce or a proof of concept,
• the impact you observed (e.g. a secret that survives redaction, a namespace that leaks across accounts, a firewall check that can be bypassed).

You can expect an acknowledgement within 5 working days and a status update at least every 14 days until the issue is resolved. We will credit you in the advisory unless you prefer otherwise.

In scope:

• Secret redaction bypasses in backend/app/firewall/secrets.py (credentials persisted to the graph, vector store, or returned by any read path).
• Cross-namespace data access (one account reading or deleting another account's memories).
• Authentication or API-key handling flaws in backend/app/identity.py / backend/app/auth.py.
• Supabase RLS or privilege mistakes in supabase/migrations/.
• Firewall bypasses that let a stale, contradicted, secret-bearing, or unsupported memory into the trusted pack.

Out of scope:

• Findings that need a compromised machine, a malicious MCP client you configured yourself, or physical access.
• Denial of service against the free hosted demo (Hugging Face Space / Vercel) by volume alone.
• Vulnerabilities in third-party services (Hugging Face, Supabase, Neo4j, Vercel) — report those upstream.

Never paste a real credential into an issue, a test fixture, or the public demo. Use obviously fake values such as the AWS documentation example key AKIAIOSFODNN7EXAMPLE.