About
For years, people predicted that "cyberwar" would take down the power grid. CyberSquirrel1 kept score and found the real threat was squirrels. Between 2013 and 2019 it logged more than 2,500 outages caused by animals, against a handful of confirmed nation-state cyber operations.
CyberAgent1 continues that tally and adds a third team. Coding agents and other autonomous AI tools now delete production databases, wipe drives and take down services. Who causes more outages: squirrels, people, or AI agents?
What counts
- Animals: any creature that caused a power, network or infrastructure outage. Same rules as CyberSquirrel1.
- Humans: cyberattacks, human error (bad config pushes, fat fingers), physical causes (cable cuts, copper theft, sabotage) and deliberate shutdowns.
- AI agents: an AI agent or coding tool acting on its own that caused an outage or destroyed production data or infrastructure. The operative is the company behind the tool (OpenAI for Codex, Anthropic for Claude Code), with the product named alongside. Disputed attributions are kept but marked in the notes.
Every entry links to a public source. When we can't find a source, we leave the incident out.
Where the data comes from
- CyberSquirrel1 dataset (2,578 animal ops, 1937–2020) by Space Rogue, with Jericho and Mathew Harper. CC BY-NC-SA 4.0.
- A hand-researched seed of notable 2019–2026 incidents, plus the confirmed incidents from CyberSquirrel1's "Human Outages" sheet.
- Status-page histories and postmortems from Cloudflare, AWS, Microsoft Azure and Google Cloud, plus major outages reported by Downdetector, each counted only when the company published a cause.
- News feeds: the GDELT Project, GDELT Cloud and Google News, plus the Rogue AI Agent Tracker ("The Datamaxxer by Eon", CC BY 4.0) and the Cloudflare Radar Outage Center (CC BY-NC 4.0). Feed items are published automatically by plain keyword rules: the headline must name an operative and an outage (for people, a stated cause), and near-duplicates of an existing entry are skipped.
- Your reports: report an op.
Why no AI reads submissions
Submitted text is untrusted, and a model that reads it can be steered by it (prompt injection). So nothing here uses AI. Public submissions sit in a queue until a human maintainer checks the source and approves or rejects them. Feed items skip the queue but pass only fixed keyword rules, so a crafted headline can at worst add one mislabeled row, which a maintainer can delete, and on the site every piece of submitted text is shown as plain text, never markup.
Caveats
- This is a count of reported incidents, not of incidents. Squirrel outages are rarely news, so the real gap is far larger than these numbers show.
- CyberSquirrel1's collection effort peaked in 2017. Years after 2018 are being backfilled.
- Locations come from the source story. When an outage was worldwide (a cloud outage, say) or the story gives no site, the entry is placed at the affected organization's headquarters (for AI agents, the company behind the tool) and labeled "Global (HQ: …)" or "HQ: …". HQ locations are best guesses.
License & credits
The dataset (download CSV) is licensed CC BY-NC-SA 4.0, the same license as the CyberSquirrel1 data it builds on. Credit: Space Rogue and Cyber Squirrel 1, with thanks to Jericho, Josh Corman (whose 2012 talk "Cyberwar: Not what we were expecting" started it) and Mathew Harper. Eon's Rogue AI Agent Tracker and Cloudflare Radar are credited above.
#CyberWar4Ever