diff --git a/core/chat_agent.py b/core/chat_agent.py
index 9c28ef7..1157059 100644
--- a/core/chat_agent.py
+++ b/core/chat_agent.py
@@ -527,6 +527,15 @@ def run_cowork(
                     preview = {"kind": "info", "title": name, "text": str(args)}
                     emit({"type": "tool_proposed", "id": tc_id, "name": name, "args": args,
                           "preview": preview})
+                    if ctx.block_network:
+                        result = {"ok": False, "output": (
+                            f"{name}: network access is blocked by the Sandbox Security Layer "
+                            '("Block network for agent-run commands" is on in Settings).')}
+                        emit({"type": "tool_result", "id": tc_id, "name": name,
+                              "ok": False, "output": result["output"]})
+                        messages.append({"role": "tool", "tool_call_id": tc_id, "name": name,
+                                         "content": result["output"]})
+                        continue
                     # R05-T04: MCP/connector tools used to run with NO permission
                     # check at all — this is what closes that gap. Same policy,
                     # same gate object as the built-in tools below.
diff --git a/core/code_agent.py b/core/code_agent.py
index 24752cb..a31cdc3 100644
--- a/core/code_agent.py
+++ b/core/code_agent.py
@@ -327,7 +327,12 @@ def run_code(
             else:
                 emit({"type": "tool_start", "id": tc_id, "name": name})
                 if is_extra and extra_executor is not None:
-                    result = extra_executor(name, args)
+                    if ctx.block_network:
+                        result = {"ok": False, "output": (
+                            f"{name}: network access is blocked by the Sandbox Security Layer "
+                            '("Block network for agent-run commands" is on in Settings).')}
+                    else:
+                        result = extra_executor(name, args)
                 else:
                     def on_output(line: str, _id=tc_id, _name=name) -> None:
                         emit({"type": "tool_output", "id": _id, "name": _name, "delta": line})
diff --git a/presentation/monitoring/tabs/tools_admin_tab.py b/presentation/monitoring/tabs/tools_admin_tab.py
index 1499c50..41858cc 100644
--- a/presentation/monitoring/tabs/tools_admin_tab.py
+++ b/presentation/monitoring/tabs/tools_admin_tab.py
@@ -216,7 +216,8 @@ class ToolsAdminTab(QWidget):
         result. Respects the fetch_url toggle: when web access is OFF the agent
         cannot reach the internet, so the test reports that instead of probing."""
         disabled = ("fetch_url" in self.ctx.config.tools_disabled
-                    or not bool(self.ctx.config.agent_security.get("allow_url_fetch", True)))
+                    or not bool(self.ctx.config.agent_security.get("allow_url_fetch", True))
+                    or bool(self.ctx.config.agent_security.get("block_network", False)))
         if disabled:
             self.test_internet_status.setText(tr("tools_admin.internet_disabled"))
             self.test_internet_status.setStyleSheet("color: #c00;")
diff --git a/state.py b/state.py
index 976be01..d1a9d6b 100644
--- a/state.py
+++ b/state.py
@@ -233,6 +233,10 @@ class AppContext:
         connections across calls/turns (spawning a subprocess per turn would
         be slow and wasteful). A server/connector that fails to connect is
         skipped, not a hard failure for the turn."""
+        # Sandbox Security Layer blocks agent-owned network connectors before
+        # they can spawn a server or issue a REST request.
+        if self.config.agent_security.get("block_network", False):
+            return [], None
         # Master switch (Monitoring → Tools → Connector): when the admin turns
         # "Connect to external" off, the agent connects to NO external
         # connectors/MCP at all — no subprocesses spawned, no REST calls.