FRO← Back to the story page · All experiments · All documentation
Under Fire: AI-Automated Security vs. a Real Botnet: full documentation
By Fractal Research Organization · Last updated Nov 8, 2026 (America/Denver). This page documents the experiment at a level that is safe to publish. Raw notebooks, settings and anything security-sensitive stay private. Everything here is early unless it says otherwise.
The question · Status and schedule · Method · What is measured · What is early, and what would count as a result · Results so far · Nulls, misses, failures and corrections · Limitations and what we do not claim · Public artifacts · Changelog
The question
A botnet started probing our server on the evening of Oct 30, 2026. We turned it into a real-world, passive test: what reaches our logs, and does our own stack (server firewall, an existing automatic ban service, FNAI Threat Watch, FNAI Sentinel, and the web server) keep up while AI agents watch?
Status and schedule
Running, ongoing. Started Oct 30, 2026. The original plan was a window through Nov 7; on Oct 31 we made it ongoing with no end date. The data refreshes once a day, around 6:15 AM Mountain Time, and the story page updates automatically.
Method
- Passive observation only. We added no new blocking for this test and we do not strike back.
- A read-only daily pull of aggregate counts from the server's firewall, SSH and web logs. Our own and agent traffic is removed.
- Outside sources are grouped by network range; individual addresses are never published.
- Web probes are grouped into generic categories. Exact paths, rules and settings are not published.
- Stack health is read from service status and from a daily outside check of the home page.
What is measured
- Outside events per hour, and the busiest hour
- Kinds of attempts and the ports knocked on most
- How many source networks, and how quiet each one is per hour
- Logins or admin access that our automated log rules flag as not ours
- Web response time (median and 95th percentile) and server errors
- Time from event to Threat Watch record
- Service restarts and the existing ban service's count
What is early, and what would count as a result
All numbers are early, passive and descriptive. They say what reached our logs, not who is behind it or why. A result we would treat as meaningful later: a sustained change in volume or tactics across weeks, or any login or admin access that a review cannot match to our own access.
Results so far
| Measure | Value (Oct 31 to Nov 8, 8 AM MT) |
|---|---|
| Outside events in our logs | 170,978 |
| Median per full hour / busiest hour | 801 / 1,027 |
| Distinct source networks | 8,037 |
| Network-hours under 45 events | 99.6% |
| Web requests / server errors | 26,550 / 0 |
| Web response median / 95th percentile | 3 ms / 13 ms |
| Automatic bans by the existing ban service | 248 |
| Accepted SSH logins, Oct 24 to Nov 8, reviewed by hand | All matched our own access (review done Nov 8) |
Earlier dated summaries are in the News archive and the daily Security Story.
Nulls, misses, failures and corrections
- Correction, Nov 8, 2026. Earlier versions of the page, the experiments card, News and our social posts said "Logins or admin access by outsiders: 0" and "nothing has got through". Those numbers came from automated log rules, and the plain wording claimed more than we had checked. On Nov 8 we reviewed every accepted SSH login on both servers from the start of our logs (Oct 24) through Nov 8; every one matched our own access. The wording now says exactly that.
- FNAI Sentinel shows a high restart count (650 at the Nov 8 pull). We have not yet explained it on the page; it means Sentinel was not running smoothly the whole time, and it is one reason we do not credit Sentinel with stopping anything.
- The web response chart has one hour with a very high 95th percentile (about 43 seconds). We have not yet explained that spike.
- An earlier version of our internal weekly summary also attributed the traffic to a single operator and stated high confidence of no compromise. We withdrew both; the public page never relied on them.
Limitations and what we do not claim
- We cannot say who runs the botnet or why. Network ranges are not identities.
- The login review covers SSH. Web-app sign-ins are counted by automated rules only.
- Whether FNAI Sentinel specifically prevented anything is unverified.
- The no-intrusion statement covers the reviewed period only; later days rely on automated rules until the next review, which we will date here.
Public artifacts
Changelog
- Oct 30, 2026: botnet probing starts; test begins.
- Oct 31: story page, card and automatic daily refresh live; window made ongoing.
- Nov 8: login wording corrected after a full review of accepted SSH logins; this documentation page added.