What this is

Global Threat Monitor is a free, ad-free, one-person project: a live map of publicly reported conflict, terrorism, cyber, and political events, plus natural disasters (earthquakes, wildfires, floods, volcanic activity, storms, drought), pulled from open news feeds and public government/scientific data sources. It refreshes automatically, keeps a rolling history you can scrub back through, and can optionally email you a digest.

It is not an intelligence product, not fact-checked by a human newsroom, and not a substitute for verified reporting on anything you're making a real decision based on. Treat it as a fast, honest first pass — then go read the original source, which is always linked.

Data sources

Every event on the site traces back to one of these. Nothing is invented — see "How events are summarized" below for exactly what role the AI plays and doesn't play.

Conflict / terrorism / cyber / political news

Natural disasters

Map & location

Every card links directly to the original article or agency record it came from — that link is resolved server-side from the actual feed item, never generated by the AI (more on that below).

How events are selected & summarized

Conflict/terrorism/cyber/political events go through a local AI model (qwen2.5:14b, running via Ollama). The pipeline is deliberately narrow:

• Raw headlines are pulled from the RSS feeds above, each tagged with its real source, timestamp, and link.
• The model is shown that numbered list and asked to pick the 6–10 most significant current security events, and to estimate coordinates from the location mentioned.
• The model writes a short title/summary and classifies category (conflict/terrorism/cyber/political) and severity (critical/high/medium).
• The model never gets to state the source, link, or timestamp itself — it can only reference a headline by its index number. The server then looks up the real link, real source name, and real publish time from that index. This is a deliberate safeguard: language models are not reliable at reproducing exact URLs or facts verbatim, so that job is never delegated to it.

Natural disaster events (earthquakes, EONET hazards) need no AI at all — they're structured data straight from USGS/NASA with real measured coordinates, magnitudes, and timestamps.

The email digest (if you subscribe) uses the same local model to write plain-language summaries of an already-machine-grouped set of events — again, only for prose, never for facts, links, or figures.

Known limitations

Being upfront about where this can be wrong or incomplete:

• Category and severity are model judgment calls, not verified facts — a story can be mis-classified or its severity over/under-stated.
• Conflict/political event coordinates are estimated by the model from the text, not GPS-verified — they're a reasonable approximation of "where," not survey-grade. Natural disaster coordinates, by contrast, come directly from USGS/NASA measurements and are precise.
• Source coverage is narrow by design — five general English-language news outlets plus ReliefWeb. Regional, non-English, and specialist sources aren't included yet, so regional stories that don't make it onto those wires won't appear here even if locally significant.
• Only the 6–10 most significant events per refresh cycle are surfaced from each batch of headlines — this is a curated significant-events feed, not a complete transcript of everything happening.
• Earthquake and EONET thresholds are fixed (currently magnitude 4.5+ for quakes) — smaller events aren't tracked.

Refresh & retention

• News feed: rechecked roughly every minute; each event stays visible for up to 7 days.
• Natural disasters: rechecked roughly every 5 minutes; same 7-day rolling window.
• Everything shown is also written to an append-only local archive that powers the history scrubber — nothing is deleted, so you can scrub back through past activity, not just what's currently on-screen.

Privacy

• No accounts, no third-party analytics, no tracking scripts, no cookies. We do keep an aggregate, privacy-preserving visit counter server-side — page loads per day, and an approximate unique-visitor count computed by hashing your IP address together with the current date. Because the date is part of the hash, it's different every single day, so it can never be linked across days or reversed back to your IP — and the raw IP address itself is never written to disk anywhere. There is no per-visitor profile, ever.
• Your watchlist and home location are stored only in your browser's local storage — they're never sent to or stored on the server.
• If you subscribe to the email digest: your email address and chosen filters are stored solely to send you that digest, confirmed via a double opt-in link, and every email includes a one-click unsubscribe that deletes your subscription immediately. Your email is never sold, shared, or used for anything else.
• Feedback submissions (with an optional email if you want a reply) are stored privately and reviewed only by the site owner — never published or shared.

Questions or something looks wrong?

Use the feedback button (💬) on the main page — I read every submission and this is a one-person project I actively maintain.