Please report vulnerabilities privately — do not open a public issue.

• Prefer GitHub's private vulnerability reporting on the repository, or
• email the maintainers directly (see the repository's contact information).

Include what you can of: affected versions/commits, a minimal synthetic reproduction (never real secrets or user data), and impact assessment. We will acknowledge private reports and coordinate fixes and disclosure with you.

Nitelos treats evidence and model output as untrusted data by design (docs/planning/07-security-and-privacy.md is the normative threat model). Hardening exists for path traversal, decompression bombs, oversized payloads and prompt-injection-style content, and errors/reprs/logs are value-free.

Out of scope by design in v0.1: Nitelos is not an adversarial process sandbox; a custom Python backend is trusted executable code; no claim is made that checks prove screenshots genuine or cryptographically attest external state. A hostile host is outside the threat model.