diff --git a/src/hooks/src/hooks/dangerous-actions/evaluate.ts b/src/hooks/src/hooks/dangerous-actions/evaluate.ts
index 097de8b9c..4c46dde8a 100644
--- a/src/hooks/src/hooks/dangerous-actions/evaluate.ts
+++ b/src/hooks/src/hooks/dangerous-actions/evaluate.ts
@@ -13,6 +13,7 @@ import {
  * Matches the `Rosetta-AI-reviewed` brand token with word boundaries on both sides.
  * Accepts any surrounding context: `# Rosetta-AI-reviewed`, `-- Rosetta-AI-reviewed`,
  * plain `Rosetta-AI-reviewed`. Rejects merged words like `XRosetta-AI-reviewedY`.
+ * Danger patterns in `patterns.ts` must never contain this token, or a match overrides itself.
  */
 const MARKER_RE = /\bRosetta-AI-reviewed\b/;
 
@@ -26,11 +27,12 @@ const MARKER_FIELDS_BY_TOOL: Readonly<Record<string, readonly string[]>> = {
   MultiEdit: ['edits'],
 };
 
-const MCP_MARKER_FIELDS = ['command', 'sql', 'query', 'new_string', 'content'] as const;
-
 const MCP_SHELL_FIELDS   = ['command', 'cmd', 'shell_command'] as const;
 const MCP_PATH_FIELDS    = ['path', 'file_path', 'filePath', 'target', 'target_path'] as const;
 const MCP_CONTENT_FIELDS = ['content', 'new_string', 'query', 'sql'] as const;
+// Markers are accepted in the fields the agent can append them to (shell and content).
+// Paths remain excluded so an override cannot be carried solely by changing an operation target.
+const MCP_MARKER_FIELDS = [...new Set([...MCP_SHELL_FIELDS, ...MCP_CONTENT_FIELDS])] as const;
 
 type PatternHit = { result: HookResult; pattern: DangerPattern | null };
 
diff --git a/src/hooks/tests/dangerous-actions.test.ts b/src/hooks/tests/dangerous-actions.test.ts
index 5f5f598d1..ab3965b83 100644
--- a/src/hooks/tests/dangerous-actions.test.ts
+++ b/src/hooks/tests/dangerous-actions.test.ts
@@ -650,6 +650,13 @@ describe('evaluateDangerous — MCP tool calls (mcp-call kind)', () => {
     expect((r as {kind:'deny';reason:string}).reason).toContain('rm-rf-root');
   });
 
+  test('mcp shell command with marker only in path → deny (paths cannot override)', () => {
+    expect(evaluateDangerous(mcpCtx(
+      'mcp__plugin_serena_serena__execute_shell_command',
+      { command: 'rm -rf /tmp/x', path: '/tmp/Rosetta-AI-reviewed' }
+    ))?.kind).toBe('deny');
+  });
+
   test('mcp filesystem write_file to .aws/credentials → advise aws-credentials (non-blocking)', () => {
     const r = evaluateDangerous(mcpCtx(
       'mcp__filesystem__write_file',
@@ -718,6 +725,34 @@ describe('evaluateDangerous — MCP tool calls (mcp-call kind)', () => {
     expect(r).toBeNull();
   });
 
+  test('mcp shell cmd with `# Rosetta-AI-reviewed` → null (marker applies to every MCP shell field)', () => {
+    expect(evaluateDangerous(mcpCtx(
+      'mcp__shell__run',
+      { cmd: 'rm -rf /tmp/x  # Rosetta-AI-reviewed' }
+    ))).toBeNull();
+  });
+
+  test('mcp shell cmd without marker → deny', () => {
+    expect(evaluateDangerous(mcpCtx(
+      'mcp__shell__run',
+      { cmd: 'rm -rf /tmp/x' }
+    ))?.kind).toBe('deny');
+  });
+
+  test('mcp shell shell_command with `# Rosetta-AI-reviewed` → null (marker applies to every MCP shell field)', () => {
+    expect(evaluateDangerous(mcpCtx(
+      'mcp__shell__run',
+      { shell_command: 'rm -rf /tmp/x  # Rosetta-AI-reviewed' }
+    ))).toBeNull();
+  });
+
+  test('mcp shell shell_command without marker → deny', () => {
+    expect(evaluateDangerous(mcpCtx(
+      'mcp__shell__run',
+      { shell_command: 'rm -rf /tmp/x' }
+    ))?.kind).toBe('deny');
+  });
+
   // Obj9: MCP marker in query field (not just command)
   test('mcp postgres query with TRUNCATE + marker in query → null (query field in MCP_MARKER_FIELDS)', () => {
     const r = evaluateDangerous(mcpCtx(