FSV — website security scanner

Web security scanner for the problems nobody is watching — dangling DNS & subdomain takeover, orphaned cloud buckets, compromised script CDNs, SPF includes that expired, debug mode left on, secrets in your JavaScript, end-of-life software, and the attack paths that turn small findings into a real compromise.

Runs as a CLI , a dashboard , an MCP server for AI agents, and a CI step .

MCP in 30 s · See it work · What it finds · Attack paths · Install · Usage · CI · FSV & ZAP · Limits · Docs · Contact

### ⚠ Authorized use only

FSV is a security assessment tool. Use it only against systems you own or are explicitly authorized in writing to test. Scanning systems without permission is illegal in most jurisdictions. Every entry point (web, CLI, MCP, CI) requires you to confirm authorization, and a shared deployment can additionally require proof of control of the target before any payload is sent (docs/DEPLOYMENT.md).

---

Connect your AI assistant in 30 seconds

Point Claude, Cursor or VS Code at FSV and ask it to review your site. The assistant asks you to confirm authorization, runs a passive check, then a full scan, leads with attack paths and hands you paste-ready fixes.

pip install fsv

fsv mcp-config --client claude-code # prints: claude mcp add fsv -- /path/to/fsv-mcp fsv mcp-config --client claude-desktop # or cursor · vscode — prints the JSON to paste

No install at all: with uv, point any MCP client at uvx fsv mcp ({ "command": "uvx", "args": ["fsv", "mcp"] }). FSV is listed in the official MCP Registry as io.github.twayst0/fsv.

Then, in the assistant: *“Use the security_review prompt on https://my-shop.example — I own it.”*

MCP tools & prompt

Prompt security_review(target, server) — the guided, authorization-first workflow above. Scan results carry an attack_paths section so the agent leads with what matters. The server is tested end to end over stdio (tests/test_mcp_stdio.py).

See it work

That output is real — a scan of the deliberately vulnerable demo site that ships in this repo. Reproduce it in a minute:

python examples/demo_site.py &                 # http://127.0.0.1:8961 — loopback only, fake secrets

fsv scan http://127.0.0.1:8961 --authorize fsv web # open the same scan in the dashboard

<sub>The `cloud_storage` check asks Amazon S3 whether the demo's bucket exists — that one request goes to AWS (read-only). Skip it with `--no-external`.</sub>

What it finds that checklists miss

Plenty of tools tell you a header is missing. The incidents that actually hurt organisations — including very well-run ones — are quieter, and they live where nobody is looking.

Details, and every false-positive guard: [docs/DETECTIONS.md](docs/DETECTIONS.md).

Attack paths

Real compromises are rarely one bug — they're two or three "medium" problems that line up. Scanners rate each in isolation, so the combination stays invisible until someone writes the incident report. FSV reads its own findings (sending nothing more) and, where a chain is actually present, adds a finding that names the links, tells the story step by step, and states the cheapest single fix that breaks it. If a link is triaged away, the chain steps aside automatically.

Dashboard

Liquid-glass UI with a live stepper (survives a page reload), the attack-path panel, OWASP breakdown, per-host risk trend, sort / module filters, bulk triage, one-click re-scan, keyboard navigation, scan comparison, and one-click export — reports or paste-ready fixes. Light and dark themes; no third-party requests.

Features

• Finds the pages first — besides links: robots.txt, sitemaps, routes leaked by debug pages, and API paths quoted in your JavaScript; action-like paths are never requested. --seed for anything else.
• A badge for your README — fsv badge writes a dated grade badge (FSV security | A · 2026-10-25); regenerate it from CI.
• Fix packs — turn findings into nginx / Apache / Caddy / Express / DNS configuration (fsv fixes nginx). Only what was actually found; CSP is emitted report-only.
• Notifications — --notify (or FSV_WEBHOOK_URL) posts the grade and the worst findings to Slack, Discord or your own endpoint. Evidence is never included.
• CI-ready — fsv.toml, baselines (gate only on *new* findings), --fail-on, SARIF for GitHub code scanning, and a ready-made GitHub Action.
• Triage that sticks — false positive / accepted risk / fixed / confirmed, stored per host by stable fingerprint, re-applied to later scans, excluded from the score. Diff any two scans; per-host trend.
• Evidence over guesses — exposure checks validate *content* (a .git/HEAD must look like one; an SPA that answers 200 to everything produces nothing). Every finding has a confidence, a CWE, an OWASP Top 10 (2021) category, impact and remediation.
• Honest failure — a crashing module is counted, explained and logged; a target that never answered is flagged, not graded "clean".
• Safe by construction — secrets masked at the source; cloud-metadata/link-local never reachable, loopback/RFC 1918 optionally blocked, on every lookup and redirect hop; dashboard CSRF/DNS-rebinding guard and token auth.
• Extensible — a module is one class; drop a file in ~/.fsv/plugins or ship an entry-point package (docs/WRITING_MODULES.md). Bring your own JS-library advisories and EOL dates.
• Fast & polite, hard to stall — async engine, shared HTTP client with cache, request coalescing, token-bucket rate limit, retries, size cap, time limit with partial results. A hanging endpoint costs one timeout per scan, not one per module; every response regex is linear-time. Tested against hostile local servers (redirect loops, 20 MB bodies, hangs, catch-all SPAs).
• Five report formats — Markdown (OWASP-grouped), JSON, self-contained HTML, SARIF 2.1, CSV.
• Local-first — results in a local SQLite file; the only optional third-party calls are listed in [DEPLOYMEN