data-breach-detector

data-breach-detector

A read-only breach-intelligence MCP server. It answers "has this organization ever been breached, what's the recent breach news, what does two decades of breach history look like, how severe is this threat text" from public disclosure feeds — and reports intelligence, not contents: the existence, timing, scale, category and exposed data-types of a breach, never the leaked records themselves.

Built for defenders and for agents that work on their behalf.

Why this instead of the alternatives

Most breach tooling sits in one of three camps, and each has a structural gap:

• Consumer checkers (HaveIBeenPwned's site) answer one question — "is my email in a breach" — one account at a time, one source at a time.
• Leak-data brokers (DeHashed, IntelX, LeakCheck and the like) sell access to the leaked records themselves. Wiring one into an AI agent hands the agent stolen credentials.
• Enterprise intel platforms (SpyCloud, Recorded Future, Flashpoint) do the join properly — behind five-figure contracts and closed APIs.

This server takes a fourth position:

1. Four primary sources, one queryable surface. The verified breach directory (HIBP), a live ransomware leak-site tracker (RansomLook), a ~16k-victim leak-site archive back to 2020 (ransomwatch), and SEC 8-K Item 1.05 filings — companies' own legally mandated "material cybersecurity incident" disclosures. Regulator-grade and criminal-infrastructure-grade evidence in the same index. No key, no contract.
2. History is first-class. breach_history, breach_timeline and breach_stats treat 2007→today as the product, not a cache: every breach of 2013, an organization's full incident chronology, repeat-victim flagging, per-year and per-actor aggregates.
3. The ethical boundary is in the code, not the terms of service. No fetch/crawl/proxy primitives, no .onion access, and every feed-authored string is sanitized where the record is built, before any field is assembled from it: emails, hashes, IPs, crypto addresses and credential-shaped tokens are redacted, and the invisible channels used to hide instructions from a human reader (Unicode Tags, zero-widths, bidi overrides, variation selectors, terminal control codes) are stripped. That matters because leak-site titles are written by ransomware crews and read by agents: the same field is both intelligence and an injection surface. Redacting only the two fields a human looks at is not enough, because ids, actor names and statistic bucket keys are built from the same strings.
4. Honesty is instrumented. feed_sources reports each feed's newest item, a staleness flag and the last fetch error — a dead upstream is a served fact, not a silent hole. (The ransomwatch project itself froze in June 2025; this server says so instead of pretending.)
5. MCP-native, free, MIT, self-hostable. One pip install, stdio or streamable-HTTP.

What it does not do

• No arbitrary URL fetch, no crawl, no proxy — no general scraping primitives.
• No .onion marketplace access, no transactions.
• Never returns the raw text of a dump, paste or leak. Feed-authored strings are redacted and stripped of hidden-instruction characters at the point each record is constructed, so ids, actor names, entity names, dates, source URLs and aggregation keys are built from sanitized values rather than raw ones.

Sources (public, no key)

• HaveIBeenPwned /api/v3/breaches — the verified breach directory back to 2007: domain, breach date, pwn count, exposed data categories.
• RansomLook (ransomlook.io) — live ransomware leak-site tracker.
• ransomwatch (joshhighet/ransomwatch) — frozen archive of ~16k leak-site posts, Jan 2020 → Jun 2025, retained as history.
• SEC EDGAR — 8-K filings carrying Item 1.05 Material Cybersecurity Incidents (mandatory first-party disclosure since Dec 2023).

Tools

tool	what it returns
breach_news(since_days, sector, source, limit, offset)	recent disclosures — entity, date, scale, exposed data types, severity
check_exposure(query, since_days, limit, offset)	does a domain/company appear anywhere in breach data — yes/no + metadata
breach_history(query, year_from, year_to, sector, data_type, min_accounts, order, limit, offset)	search the full archive back to 2007
breach_timeline(entity, limit, offset)	one organization's incident-by-incident chronology + repeat-victim assessment
breach_stats(group_by, sector, limit)	aggregates per year / source / data type / threat level / ransomware actor
assess_threat(text)	classify a piece of security text — level, categories, action (no network)
feed_sources()	feeds, per-source freshness, staleness flags, last fetch errors

Every list tool reports count, limit, offset and returned, and breach_stats reports buckets_total, so a truncated answer is visible as truncated and the tail is reachable by paging rather than lost.

Run

pip install data-breach-detector

data-breach-detector           # stdio (for MCP clients)
data-breach-detector --http    # streamable-HTTP on 127.0.0.1:8790/mcp

Or point an MCP client at the config:

{ "mcpServers": { "data_breach_detector": {
  "command": "data-breach-detector"
} } }

Hosted remote: https://breach.seiche.info/mcp

License

MIT. The breach data belongs to its sources (HaveIBeenPwned, RansomLook, ransomwatch, SEC EDGAR); this tool only aggregates their public disclosure metadata, with attribution.