+
+
+ {/* Policy fix */}
+
+ Suggested Fix — Tool Policy
+
+ }
+ headStyle={{ borderBottom: "1px solid #e0e7ff", backgroundColor: "#f5f3ff" }}
+ >
+ {policyApplied ? (
+ }
+ message="Policy applied! bash_tool and code_execution are now blocked for Claims Agent."
+ description="The agent has been re-tested. Tool calls return a policy-violation error before reaching the LLM. You can verify by re-running a sample request below."
+ className="rounded-lg"
+ />
+ ) : (
+
+
+ The following tools were detected outside the Claims Agent's permitted scope. Select the tools to block and
+ apply the policy to prevent future violations.
+
+ The semantic similarity between this agent's description ("answer member questions about insurance claims") and the tools it is calling (exec, write, fetch_claims) is very low. The gateway embeds both on every run and computes cosine similarity — this agent is scoring consistently near zero.
+
+
+
+ Similarity — description vs tools called
+ 0.08
+
+
+ Expected range for this agent type
+ 0.82 – 0.95
+
+
+
+
+
+ {FETCH_CLAIMS_INJECTION}
+
+ {/* Secondary stats bar — small, below the alert */}
+
+ {agent.totalCalls.toLocaleString()} total calls
+ {agent.blockedCalls} flagged calls
+ {agent.untrustedTools} unlisted tools
+ {agent.maxIterations} max iterations
+
+
+
+ {/* ── ACT 2: Why it happened ── */}
+
+
Why it happened
+
+ {/* Agent Profile */}
+
+
Agent Profile
+
Claims Processing Agent
+
"Answer member questions about insurance claims, coverage, and billing"