Week 46, 2026
9–15 November 2026 (in progress, updated live)
158 visits to the MCP server and 210 to the web traps; 4 attackers, 151 leak scanners, 156 registry crawlers; 8 MCP visits called a tool; 52 fake secrets handed out, 0 reused.
Totals
so far- MCP visits
- 1581,504 MCP messages
- Web visits
- 2101,525 web requests
- Secrets handed out
- 52
- Secrets reused
- 0
- Followed the server's instructions
- 3
Hour by hour
requests, UTCMCPpeak 100Webpeak 670
Day by day as a table
| Day (UTC) | MCP messages | Web requests |
|---|---|---|
| Mon 9 Nov | 1,413 | 1,337 |
| Tue 10 Nov | 91 | 188 |
| Wed 11 Nov | 0 | 0 |
| Thu 12 Nov | 0 | 0 |
| Fri 13 Nov | 0 | 0 |
| Sat 14 Nov | 0 | 0 |
| Sun 15 Nov | 0 | 0 |
Key findings
so far- 112 of 158 MCP visits read the tool list; 8 (5%) went on to call a tool.
- Leaving out directory, registry and health checkers, 71 MCP visits remain and 2 of them called a tool.
- 3 visits followed the server's instructions and called get_started first.
- The commonest behaviour code was R2 C0 U0 X0 I0 A0 (102 visits): recon: Enumerated.
- The most-called tool was get_started (3 times).
- The most common MCP client name was AgentsJunction health checker (6 times).
- Web scanners looked hardest for source control (760 times).
- As a normal site would see it, 16% of web visits were unwanted (leak hunting, attacks, robots.txt ignored or a faked crawler name) and 20% were wanted.
- 81% of MCP visits were wanted: clients and directories connecting, and agents using the tools.
- Judged by the gaps between their requests, 88% of MCP visits that could be timed ran like scripts (machine speed or a fixed wait), 13% paused like a model and 0% like a person.
- 8 MCP client names came back on a schedule; the busiest was mcpbeat, every 16 minutes.
- Most visits came from United States (151).
- None of the 52 fake secrets handed out was reused.
How far they got
visits per stageWho came
by what each visit didMost advanced visitors
| Visitor | Network | Sorted as | Got to |
|---|---|---|---|
Browser | TC DATACENTER LIMITED MANAGED IP SPACE, Netherlands | Attacker | Took a secret |
Browser | HOS-2475878, Germany | Scanner | Took a secret |
Browser | Oracle Public Cloud, United States | Scanner | Took a secret |
Browser | WEDEPLOY SAS, Hong Kong | Scanner | Took a secret |
BrickBlueBot | CLODO CLOUD SERVICE CO. L.L.C, Netherlands | Explorer | Took a secret |
Unrecognised software | A100 ROW Inc, United Kingdom | Scanner | Took a secret |
Security scanner | DigitalOcean, LLC, United Kingdom | Scanner | Took a secret |
mcp-runtime-engine-auth-probe | Robi Axiata Limited - Data, Bangladesh | Registry | Took a secret |
MCP visits by level
behaviour codes v1| Dimension | None | 1 | 2 | 3 |
|---|---|---|---|---|
| Recon | 8% | 22% | 66% | 4% |
| Credential access | 98% | 0% | 1% | 1% |
| Credential use | 100% | 0% | 0% | 0% |
| Exploitation | 100% | 0% | 0% | 0% |
| Impact | 100% | 0% | 0% | 0% |
| Agent compliance | 98% | 2% | 0% | 0% |
Web visits by level
behaviour codes v1| Dimension | None | 1 | 2 | 3 |
|---|---|---|---|---|
| Recon | 40% | 10% | 47% | 4% |
| Credential access | 93% | 0% | 7% | 0% |
| Credential use | 99% | 1% | 0% | 0% |
| Exploitation | 100% | 0% | 0% | 0% |
| Impact | 100% | 0% | 0% | 0% |
Each visit scored 0 to 3 on the behaviour codebook.
MCP visits, wanted or not
As a normal site would see it. How it is judged
Web visits, wanted or not
As a normal site would see it. How it is judged
MCP visits by pace
3 not judged yet.
By the gaps between requests, not the name. How it is judged
Web visits by pace
5 not judged yet.
By the gaps between requests, not the name. How it is judged
On a schedule
MCP clients whose visits came back at regular times, with how many visits did.