AiStackGuru · Field tool 02

What one bad instruction reaches

Assume an agent will, at some point, faithfully execute an instruction it read somewhere. The question isn't whether that happens. It's what it can reach when it does. Permissions compose — and the dangerous configurations are almost always horizontal.

What has this agent been granted?

Tick what's true of one agent — not your whole stack. Run it again per agent. Nothing here leaves your browser.

None of the mitigations depend on the model behaving. That's the point of them. Filtering and better instructions reduce volume; they aren't boundaries, because they need to succeed every time against input someone else chose.

Watch the boundary-layer build

A scoring heuristic, not an assessment of your specific system — use it to find the combination you hadn't considered, not as a verdict. Anything touching financial accounts deserves a real review; guidance here is educational and is not financial advice.