CADIS will execute tools, edit files, call models, and coordinate agents. Security is a core product requirement, not a later hardening step.

Use GitHub Security Advisories. File a private vulnerability report at:

https://github.com/Growth-Circle/cadis/security/advisories/new

Private vulnerability reporting is enabled on this repository. Do not open public issues for security vulnerabilities.

We aim to acknowledge reports within 48 hours and provide an initial assessment within 7 days. Critical issues affecting approval bypass, command execution, or credential leakage will be prioritized for expedited fixes.

Report categories include but are not limited to:

• approval bypass
• command execution
• filesystem escape
• credential leakage
• prompt injection leading to tool misuse
• unsafe network access
• log redaction failure
• sandbox failure

• Risky actions require central approval.
• Secrets must never be logged.
• OpenAI API keys must come from CADIS_OPENAI_API_KEY or OPENAI_API_KEY, never from committed config, examples, or logs.
• ChatGPT/Codex account auth must stay with the official Codex CLI. Treat ~/.codex/auth.json as password-equivalent and never commit, copy, or log it.
• Tool calls must declare risk class and workspace boundary.
• Approval resolution must be atomic and first-response-wins.
• Shell execution must be traceable, cancellable, and policy-gated.
• Persistent logs should use redaction before write.

If a credential, token, event log, diagnostic bundle, or crash dump is committed:

1. Revoke or rotate the affected credential immediately.
2. Remove the file from the index with git rm --cached when it should stay local.
3. Rewrite public history when the secret reached a shared remote.
4. Invalidate provider, Telegram, or integration tokens that may have been exposed.
5. Add or tighten an ignore pattern, redaction rule, or secret scan before continuing.