Skip to content

Security: psi16181918161phi/engineering-studio-ai

SECURITY.md

title Security Policy — Engineering Studio AI
author Hadrian Hu
date 2026-07-07
version 2026.0.1.0
keywords
security
vulnerability-disclosure
secrets
hackathon

Security Policy — Engineering Studio AI

Governing standards: AGENTS.md §8 (condensed security baseline) and .github/agents/STANDARDS_SUMMARY.md (extended coverage). This repo is a public spin-off of a private standards corpus — see AGENTS.md's intro for why the full corpus is not reproduced here.


Supported Versions

This is a short-lived hackathon submission (AMD LabLabAI, Act II — Unicorn Track), not a long-term maintained product. There is a single active line of development:

Component Supported Notes
main branch (current release) ✅ Yes Active during and shortly after the hackathon
Any tagged/archived snapshot ❌ No Point-in-time submission artifact only

Reporting a Vulnerability

Do NOT open a public GitHub issue for security vulnerabilities.

  1. Open a private security advisory on GitHub, or contact a maintainer directly if advisories are unavailable.
  2. Include: affected file(s), a description of the vulnerability (CWE reference if known), reproduction steps, and potential impact.

Response

Given the hackathon timeline, response is best-effort, not SLA-backed:

Milestone Target
Acknowledgement 3–5 calendar days
Fix or mitigation As soon as practical; critical findings prioritized immediately

Scope

In scope:

  • src/engineering_studio/ — API/SSE routes (api/), the FastAPI app (webapp/), CLI (cli/), SDK (sdk/), orchestrator/agents (agents/) — injection, path traversal, insecure file handling, secrets handling.
  • frontend/ — served static assets, CORS configuration.
  • .github/workflows/ci.yml — secrets leakage, supply-chain risk.
  • Dependency set in pyproject.toml/requirements*.txt (tracked via pip-audit in CI).

Out of scope:

  • Denial of service against local dev tooling.
  • Findings in research/, docs/, paper/, presentation/ (no code execution).
  • The Fireworks AI hosted models themselves (report to Fireworks AI, not here).

Secrets Handling

  • FIREWORKS_API_KEY and any other credential live only in a local .env (gitignored) — never in .env.example, source, tests, or fixtures.
  • If a secret is ever accidentally committed: rotate it immediately at the provider, then open a private security advisory — do not rely on git rm/history rewrite alone, since the key must be treated as compromised the moment it was pushed.

Security Standards Reference

  • AGENTS.md §8 — no secrets in source, input validation/sanitization before prompt interpolation, pinned dependencies, pip-audit before submission commits.
  • .github/agents/STANDARDS_SUMMARY.md — extended coverage (OWASP- aligned baseline, CI security gates).
  • Local quality gate (ruff, mypy --strict, bandit -ll, pip-audit) — see CONTRIBUTION.md §7; all four run in .github/workflows/ci.yml.

Changelog

Version Date Author Description
2026.0.1.0 2026-07-07 Hadrian Hu Initial security policy.

There aren't any published security advisories