# Security Policy ## Supported versions | Version | Supported | |---|---| | Phocinae-Largha-150M-v1 (model weights) | ✅ | | phocinae-server / phocinae-guard / phocinae-mcp (PyPI) | ✅ latest only | | dsh-phocinae (npm) | ✅ latest only | ## Reporting a vulnerability Phocinae-Largha is a local inference model and toolset. Security-relevant concerns include (but are not limited to): prompt-injection behaviour of the decision heads, calibration exploitation, gate bypass scenarios in `phocinae-guard`/`dsh-phocinae`, and supply-chain issues in released artifacts. **Please do NOT open a public issue for vulnerabilities.** Instead: 1. Report via the private channel listed in the GitHub Security tab of this repository ("Report a vulnerability"), or 2. Email a description (scope, reproduction steps, impact) to the maintainer via the address in the GitHub profile of the repository owner. What to include: affected component and version, steps to reproduce, expected vs actual behaviour, and (if available) a patch suggestion. ## What to expect - Acknowledgment within 7 days. - Status update within 30 days, including whether the report is accepted. - Disclosure: after a fix is released, we credit reporters (or keep them anonymous on request) in the release notes / CHANGELOG. ## Scope notes - The model itself is a research artifact; it makes probabilistic decisions. It is NOT a security boundary on its own — `phocinae-guard` and `dsh-phocinae` default to fail-closed (deny) precisely because of this. - Known limitations (e.g., knowledge-heavy reasoning) are documented in `docs/technical-report.md` and `MODEL_CARD.md`; reports duplicating documented limitations may be closed as known issues.