Security tooling and AI agents are only as trustworthy as the environment they run in. This page describes the controls around my lab. It describes design and verified checks, not a claim of formal certification.
Isolation
An assessment console runs in a disposable virtual machine with hardware virtualization, on a private network with no bridge to my home LAN.
Host firewall rules reject all outbound traffic from the VM by default. Provisioning access was granted only temporarily and then removed.
The console is reachable only on the host loopback interface. No host directories or credentials are shared with the guest.
Cost and scope guards
Model calls to a paid provider pass through a local gateway that pins the model, caps output, counts input before admission, and reserves cost against a fixed allowance with a persistent ledger that is never reset.
Target traffic goes through a restricted helper: exact hostnames, HTTPS only, GET only, public IP checks, rate limit, request and time caps.
The gateway is stopped between runs so future use fails closed.
Documentation per run
Each run keeps an authorization record, an exact scope file, a run window, redacted request/response evidence, a ledger of requests and spend, and a report that states limitations and what failed. A run that stopped at a budget boundary is reported as stopped, not as successful.
Benchmarks (in progress)
I am preparing a scored comparison of open tooling against a deliberately vulnerable practice application that I host myself. Results, including false positives, time and cost, will be published here when complete. Nothing is published yet.
weneedthebeans is a brand of A Bean Consult, LLC, an Alabama limited liability company.