Skip to content

Repository navigation

Security: vinsoc-cyber/VulnSight

Security

Report a vulnerability

SECURITY.md

Security policy

VulnSight is a local research UI for querying GHSA, NVD, and OSV by bug class. It is not a multi-user SaaS. The process holds the operator's AI keys, GitHub token, and optional NVD key, and will spend them on /api/search and /api/ai/*.

The supported deployment is loopback-only (http://127.0.0.1:5000) with credentials in a git-ignored .env. Anything beyond that is an explicit opt-in and needs a shared secret.

Supported versions

There are no numbered releases. Security fixes land on main and apply to the latest commit only. Older snapshots are unsupported.

Reporting a vulnerability

Do not open a public GitHub issue for an exploitable finding (auth bypass, CSRF that spends keys, secret leakage, RCE, SSRF, and similar).

  1. Use GitHub private vulnerability reporting on this repository.
  2. If that form is unavailable, email the maintainer at the address on sondt99's GitHub profile. Put VulnSight in the subject.

Include:

  • affected commit SHA (or main as of a date)
  • impact (who can exploit it, and what they gain)
  • reproduction steps or a proof of concept
  • whether you have already disclosed it elsewhere

You should get an acknowledgement within 7 days, and a decision (fix / won't-fix / not a vulnerability) within 30 days. Please give at least 90 days after a fix is on main before public write-ups, unless we agree otherwise.

Public issues are fine for hardening ideas, docs gaps, and non-exploitable bugs.

Threat model

In scope Out of scope
Unauthenticated or CSRF use of /api/search, /api/ai/test, /api/ai/classify that spends API keys or GitHub credentials A local operator attacking their own loopback instance
Secret leakage in HTTP responses, HTML, logs shipped to clients, or committed files Prompt injection that only skews AI verdicts (advisory text is untrusted evidence by design)
XSS, HTML injection, or CSV formula injection in the UI / export Findings that require --debug, a world-readable .env the operator created, or a compromised gh login
Path traversal, zip-slip, SSRF, SQLi, command injection in this codebase Vulnerabilities in GHSA / NVD / OSV / the configured AI provider
Bind/auth bypass that exposes the UI beyond loopback without VULNSIGHT_API_TOKEN Dependency CVEs with no reachable path in VulnSight (report them upstream too)

Operator requirements

Follow these if you load real credentials:

  • Keep .env out of git (it is git-ignored). Never commit GLM_TOKEN, AI_TOKEN, ANTHROPIC_TOKEN, GH_TOKEN, NVD_API_KEY, or VULNSIGHT_API_TOKEN. ./run.sh sets .env to mode 0600.
  • Prefer ./run.sh or docker compose as documented. Compose publishes 127.0.0.1:5000 only.
  • Do not run docker run -p 5000:5000 … unless you intend to expose the UI. Gunicorn listens on 0.0.0.0 inside the container. If credentials are loaded and VULNSIGHT_API_TOKEN is unset, the app generates a token, saves it as .vulnsight_api_token in the data dir, and requires it on POST /api/*.
  • If you bind beyond loopback (HOST=0.0.0.0, a LAN compose port, a reverse proxy): set a long random VULNSIGHT_API_TOKEN (or accept the generated one), set VULNSIGHT_PUBLIC_HOST to the DNS name the browser uses, and do not use python app.py --debug.
  • python app.py refuses a non-loopback bind while credentials are loaded unless both VULNSIGHT_EXPOSE=1 and VULNSIGHT_API_TOKEN are set, and refuses --debug unless the bind is loopback.

Relevant knobs (see .env.example):

Variable Role
VULNSIGHT_API_TOKEN Required on POST /api/* (X-VulnSight-Token or Authorization: Bearer)
VULNSIGHT_EXPOSE Opt-in for python app.py on a non-loopback HOST
VULNSIGHT_PUBLIC_HOST Extra CSRF Origin hosts (comma-separated DNS names)
VULNSIGHT_SEARCH_RATE / VULNSIGHT_AI_RATE Per-IP limits (default 30 / 20 per 60s)

What the app already enforces

  • JSON Content-Type on mutating APIs (rejects simple form CSRF).
  • Origin/Referer parsed with urlsplit (not a prefix match); Sec-Fetch-Site: cross-site is blocked.
  • Generic client errors for AI provider HTTP bodies and gh stderr; raw detail stays in server logs.
  • CSP (script-src 'self' + per-response nonce), X-Frame-Options: DENY, X-Content-Type-Options: nosniff.
  • CSV export prefixes formula-like cells with '.

Secrets and data

Item Handling
.env Git-ignored. Placeholders only in .env.example.
advisories.db / osv_cache/ Local cache of public advisory data plus AI verdicts. Git-ignored. Not a secrets store, but do not publish the files if they contain internal notes you added.
UI token prompt VULNSIGHT_API_TOKEN is never rendered into HTML; the tab stores it in sessionStorage only.

Safe harbor

Good-faith research against your own checkout, or a loopback instance you control, is welcome. Do not use findings to access other people's keys, GitHub accounts, or AI quotas.

There aren't any published security advisories