fix(agent): taint model-visible tool responses

2026-09-18 22:22:21 +02:00 · 2026-08-15 07:18:25 +00:00
parent 7a138e8a3f
commit 94cf119b11
3 changed files with 141 additions and 21 deletions
@@ -3060,15 +3060,19 @@ def _append_tool_results(
                "content": result_text,
            }
            capabilities = capabilities_for_action(tool_name, tool_content)
-            if capabilities.result_integrity is not ResultIntegrity.SYSTEM:
+            should_arm_gate = tool_result_should_arm_gate(
+                tool_name,
+                result,
+                tool_content,
+            )
+            if (
+                capabilities.result_integrity is not ResultIntegrity.SYSTEM
+                or should_arm_gate
+            ):
                result_message["metadata"] = {
                    "trusted": False,
                    "source": f"tool result: {tool_name}",
-                    "tool_gate_untrusted": tool_result_should_arm_gate(
-                        tool_name,
-                        result,
-                        tool_content,
-                    ),
+                    "tool_gate_untrusted": should_arm_gate,
                }
            messages.append(result_message)
    else: