The two choices in the console
A third mode, your own sandbox, exists for agents DeepBlocker set up with you. You cannot pick it in the add form.
Connect our test tools
You make a test copy of your agent. In that copy you swap its real tools for ours. Our tools answer with safe, made-up data and record every call. Your live agent is never touched. With this mode we can show a finding at any level, because the proof is a tool call we served ourselves. See Connect our test tools.Real tools still connected
The add form has a box: I removed my agent’s own real tools from the test copy. Until you tick it, we treat the real tools as still connected. We then use only attacks that ask your agent for words, never for actions, so nothing real can happen. Tick the box once the test copy has only our tools.Test it as it is
We reach the agent on its real channel and judge each reply from the words alone. This is the mode for a Telegram bot, or for a web agent you cannot copy. A finding from words alone is capped at Moderate. The agent may have said it would do something; we did not see it happen.Activity logs
Whichever mode you pick, your agent can send us its own activity logs over OpenTelemetry. A tool call seen in your logs grades higher than words alone and lower than our own tools: capped at High. See Send activity logs.Changing the mode
Open the agent’s Connection tab, find Change this agent, pick the other mode and press Save changes. The mode cannot change while a test of the agent is scheduled, running or paused. Stop the test first.Where to go next
Connect our test tools
The MCP address, the key and the set-up per framework.
Send activity logs
OpenTelemetry endpoints and settings.
Reports
How levels and caps become a band.