High-risk tools in Deliberation
8 of the 11 tools in Deliberation are classified as high risk. This page profiles those tools specifically, with recommended policy actions and the attack patterns that target them.
Every operation listed below is an action PolicyLayer recommends controlling at the transport layer. Open any tool to see the full profile, risk score, and YAML policy snippet.
Tools at high risk
-
consensusExecuteRun the FULL multi-round consensus convergence loop server-side with a provider arbiter (blind pass + peer fan-out -> adjudicate -> revise) and return the converged verdict. Def...
-
consensus-stepExecuteClient-driven consensus loop where YOU (the host model) are the arbiter, one action per call: init (returns sessionId + blind prompt) -> record_blind (your pre-commit verdict) -...
-
geminiExecuteStart a new Gemini expert session
-
gemini-replyExecuteContinue an existing Gemini session
-
grokExecuteStart a new Grok (xAI) expert session. Advisory only (no filesystem editing). Supports attaching files.
-
grok-replyExecuteContinue an existing Grok session (in-memory; lost if the MCP server restarts).
-
openrouterExecuteStart an OpenRouter expert session (advisory only). Pick a configured
-
openrouter-replyExecuteContinue an OpenRouter session by threadId (in-memory; lost on restart).
Attacks that target this class
High-risk tools in any server share these documented attack patterns. Each links to the full case and the defensive policy.