| Version | Supported |
|---|---|
| 2.37.x | Yes |
| 2.36.x | Yes |
| < 2.36 | No |
Do not report security vulnerabilities via public GitHub issues.
Email nick@pueo.io with:
- Description of the vulnerability
- Steps to reproduce
- Potential impact
- Suggested fix (if any)
You should receive a response within 72 hours. Critical issues will be patched and released as soon as possible.
Deepr includes:
- Input validation and request sanitization
- Prompt-injection detection/sanitization (
utils/prompt_security.PromptSanitizer) - SSRF protection on user-supplied URLs
- Scheduled fleet heartbeat endpoints require public HTTPS, reject credentials and fragments, connect only to a prevalidated public address with environment proxies disabled, refuse redirects, and never log the secret-bearing target
- API key redaction in logs and error output
- Budget enforcement to prevent runaway spend (per-operation + daily/monthly caps, approval tiers)
- Dependency audit (
pip-audit, blocking in CI) and SBOM generation - Pre-commit hooks (ruff lint, debug statement detection)
- Comprehensive automated test suite with an enforced coverage gate (see ROADMAP for current counts)
Scope note: Deepr orchestrates bounded hosted model APIs, local Ollama, and explicit plan-quota CLI backends. It does not train, fine-tune, or serve model weights. Its security focus is ingested and untrusted data plus agentic tool use (prompt injection, tool abuse, and trust boundaries), not model-internals attacks (poisoning, weight extraction, and similar provider concerns). See the "AI/agentic security" subsection of ROADMAP for the planned hardening and explicit non-goals.
See docs/ARCHITECTURE.md for technical details.