Ran this week
Example Recon Skill at commit 3f2a9c1 (v0.4.2), Claude Code in mode dontAsk, model claude-sonnet (dated 2026-09-01), 35:40 minutes, $4.37 measured at the gateway. Target: OWASP Juice Shop in the lab.
Declared vs. observed
The skill declares network to its target (crt.sh), nmap and narrow pre-approvals. In the lab it went beyond that on three of six rights: it wrote ~/recon-reports/target.md outside the working directory, tried to reach api.example.net (blocked) and ran nmap through a shell. The one that counts: the third-party API call appears nowhere in the README.
Verdict and why
So what for your agent policy
One link worth your time
A placeholder for one external article with a short comment, the author named and quoted with context.
Catalog delta
- New: Example SIEM Query (listed)
- Tried: Example Vuln MCP (beyond declared on 1 of 6 rights)