NOODUM is experimental software (0.x). It has not been audited by a third party. Do not run it with real users or real personal data yet.

Please report privately — do not open a public issue.

• GitHub: open a private security advisory
• Response target: acknowledgement within 5 business days

Include: what you found, how to reproduce it, and the impact you believe it has. If you have a suggested fix, even better.

In scope:

• The AI governance boundaries (custom/modules/aiops/components/Governance.php) — in particular anything that lets a level-2 or level-3 capability be executed autonomously.
• Prompt injection that survives all three barriers and results in an action, not merely a wrong classification.
• Audit trail tampering or omission.
• Containment that does not expire, or that cannot be reverted.
• Secret leakage into logs, rendered pages or the database.

Out of scope:

• Vulnerabilities in HumHub itself → report to https://github.com/humhub/humhub
• The demo compose.yml being insecure when exposed to the internet. It is documented as local/demo only.
• Missing email configuration. That is a known limitation, documented in the README.

All user content is untrusted data and never becomes instruction. The model may only classify into a closed label set and summarise already-computed numbers; it cannot request an action. Irreversible capabilities have no execution path in the code at all. Unmapped capabilities default to human-only. If the model provider is unavailable, the layer degrades to deterministic rules and no moderation permission fails open.