Likely blockers
The missing controls and unstable configurations that would make the planned result unreliable.
AgentBoundary
A test result means very little if the prompts, tools or permissions changed halfway through the run. Before an AgentBoundary assessment, we review the proposed configuration, the approval controls and the logging, and tell you what has to change first.
Included in the work
01
02
The missing controls and unstable configurations that would make the planned result unreliable.
The prompts, model configuration, memory behavior, tool permissions and approval rules the test will use, written down, plus a log of anything that changes during the assessment.
Go to formal testing, widen to a broader assessment, or fix some things first.
03
The open AgentBoundary benchmark sends 51 prompt-injection scenarios through chat-completion APIs and looks at how the model responds. It doesn't execute tools in a live application. If a commercial assessment also checks application permissions or downstream actions, those are separate tests and the report has to say so.
Limits
Related services and resources
Next step