Injectionator build notes
Probe and MCP: the evidence trail
This guide restores the destination referenced by the jq article. Prepared on 5 November 2026 from the published accounts below, it is not a reconstructed February release log. The original dated build record is not available here.
Follow the application action
A probe is a controlled attempt to exercise an application path. For an agent that can call tools, the useful question extends beyond its final answer: which tool did it request, under whose identity, with what arguments, and what actually happened? An MCP connection is one way to expose tools; the presence of that connection does not establish authorization or prove a control worked.
Read the report in layers
- Record the scenario and intended outcome before running it.
- Keep the input, model response, requested tool calls and observed effects together.
- Distinguish a requested action from a completed downstream effect.
- Record which control made the decision. A blocked result without attribution cannot establish inspector coverage.
- Repeat the same scenario after a correction, preserving the earlier result for comparison.
What the surviving records support
The jq walkthrough supplies concrete report-query examples. Use its examples against the actual schema of your report; field names are not a permanent API contract.
The inspector attribution account explains the coverage rule and why an apparently successful demo was insufficient evidence.
Neither account supplies a reproducible February release manifest, a current installation contract or independent production results. For current tool instructions, use the AI Surface Researcher guide.