diff --git a/apps/api/src/plugins/mcp/providers/argocd_provider.py b/apps/api/src/plugins/mcp/providers/argocd_provider.py
index 9bb6495b..ca800469 100644
--- a/apps/api/src/plugins/mcp/providers/argocd_provider.py
+++ b/apps/api/src/plugins/mcp/providers/argocd_provider.py
@@ -33,6 +33,12 @@ _HTTP_TIMEOUT = 10.0
 _RE_SAFE_APP_NAME = re.compile(r'^[a-zA-Z0-9][a-zA-Z0-9._-]{0,62}$')
 
 
+def _tls_verify() -> bool:
+    """P2 fix 2026-04-11: 讀取 ARGOCD_VERIFY_TLS env，預設 False（因 self-signed cert）。"""
+    import os
+    return os.environ.get("ARGOCD_VERIFY_TLS", "false").lower() in ("true", "1", "yes")
+
+
 def _validate_app_name(name: str) -> str:
     if not _RE_SAFE_APP_NAME.match(name):
         raise ValueError(f"Unsafe app name: {name!r}")
@@ -152,7 +158,7 @@ class ArgoCDProvider(MCPToolProvider):
         if ns_filter:
             params["appNamespace"] = ns_filter
 
-        async with httpx.AsyncClient(timeout=_HTTP_TIMEOUT, verify=False) as client:
+        async with httpx.AsyncClient(timeout=_HTTP_TIMEOUT, verify=_tls_verify()) as client:
             resp = await client.get(url, headers=self._headers(), params=params)
             resp.raise_for_status()
             data = resp.json()
@@ -183,7 +189,7 @@ class ArgoCDProvider(MCPToolProvider):
         app_name = _validate_app_name(parameters["app_name"])
         url = f"{self._base_url()}/api/v1/applications/{app_name}"
 
-        async with httpx.AsyncClient(timeout=_HTTP_TIMEOUT, verify=False) as client:
+        async with httpx.AsyncClient(timeout=_HTTP_TIMEOUT, verify=_tls_verify()) as client:
             resp = await client.get(url, headers=self._headers())
             resp.raise_for_status()
             app = resp.json()
@@ -229,7 +235,7 @@ class ArgoCDProvider(MCPToolProvider):
 
         # ArgoCD history endpoint
         history_url = f"{self._base_url()}/api/v1/applications/{app_name}"
-        async with httpx.AsyncClient(timeout=_HTTP_TIMEOUT, verify=False) as client:
+        async with httpx.AsyncClient(timeout=_HTTP_TIMEOUT, verify=_tls_verify()) as client:
             resp = await client.get(history_url, headers=self._headers())
             resp.raise_for_status()
             app = resp.json()
@@ -255,7 +261,7 @@ class ArgoCDProvider(MCPToolProvider):
     async def health_check(self) -> bool:
         try:
             url = f"{self._base_url()}/api/v1/applicationsets"
-            async with httpx.AsyncClient(timeout=5.0, verify=False) as client:
+            async with httpx.AsyncClient(timeout=5.0, verify=_tls_verify()) as client:
                 resp = await client.get(url, headers=self._headers())
                 return resp.status_code < 500
         except Exception:
diff --git a/apps/api/src/plugins/mcp/providers/prometheus_provider.py b/apps/api/src/plugins/mcp/providers/prometheus_provider.py
index ad28a18d..e2636a92 100644
--- a/apps/api/src/plugins/mcp/providers/prometheus_provider.py
+++ b/apps/api/src/plugins/mcp/providers/prometheus_provider.py
@@ -19,6 +19,7 @@ Prometheus MCP Tool Provider — MCP Phase 2b
 @see docs/superpowers/specs/2026-04-10-infra-rebuild-sprint-abc-design.md §MCP-2b
 """
 
+import re
 import uuid
 from datetime import UTC, datetime, timedelta
 from typing import Any
@@ -26,6 +27,9 @@ from typing import Any
 import httpx
 import structlog
 
+# P1 fix 2026-04-11: alertname 白名單防止 PromQL label injection
+_RE_SAFE_ALERTNAME = re.compile(r"^[a-zA-Z0-9_-]{1,64}$")
+
 from src.plugins.mcp.interfaces import MCPTool, MCPToolProvider, MCPToolResult
 
 logger = structlog.get_logger(__name__)
@@ -248,6 +252,9 @@ class PrometheusProvider(MCPToolProvider):
 
     async def _alert_history(self, params: dict) -> dict:
         alertname = params["alertname"]
+        # P1 fix 2026-04-11: 白名單驗證防止 PromQL label injection
+        if not _RE_SAFE_ALERTNAME.match(alertname):
+            raise ValueError(f"Unsafe alertname: {alertname!r}")
         window_hours = int(params.get("window_hours", 24))
         limit = int(params.get("limit", 20))
 
diff --git a/apps/api/src/plugins/mcp/providers/sentry_provider.py b/apps/api/src/plugins/mcp/providers/sentry_provider.py
index 4edf3182..21535ad8 100644
--- a/apps/api/src/plugins/mcp/providers/sentry_provider.py
+++ b/apps/api/src/plugins/mcp/providers/sentry_provider.py
@@ -31,6 +31,11 @@ logger = structlog.get_logger(__name__)
 
 _HTTP_TIMEOUT = 10.0
 _RE_SAFE_ISSUE_ID = re.compile(r'^\d{1,20}$')
+# P2 fix 2026-04-11: Sentry query 語意白名單 — 只允許已知安全的 Sentry 搜尋語法
+# 允許: 純文字識別字、is:unresolved、level:error、project:xxx、assigned:me 等
+_RE_SAFE_SENTRY_QUERY = re.compile(
+    r'^[\w\s:.\-/]+$'  # alphanumeric, space, colon, dot, dash, slash — Sentry filter 所需
+)
 
 
 class SentryProvider(MCPToolProvider):
@@ -222,7 +227,10 @@ class SentryProvider(MCPToolProvider):
         return MCPToolResult(success=True, data=result)
 
     async def _search_issues(self, parameters: dict) -> MCPToolResult:
-        query = str(parameters["query"])[:200]  # 限制長度防止注入
+        query = str(parameters["query"])[:200]
+        # P2 fix 2026-04-11: 語意白名單驗證，拒絕含特殊字元的 query
+        if not _RE_SAFE_SENTRY_QUERY.match(query):
+            return MCPToolResult(success=False, error=f"Unsafe sentry query: {query!r}")
         limit = max(1, min(int(parameters.get("limit", 10)), 25))
 
         url = f"{self._base_url()}/api/0/projects/{self._org()}/{self._project()}/issues/"
diff --git a/apps/api/src/plugins/mcp/providers/ssh_provider.py b/apps/api/src/plugins/mcp/providers/ssh_provider.py
index bfe3b9a3..2bcab0e4 100644
--- a/apps/api/src/plugins/mcp/providers/ssh_provider.py
+++ b/apps/api/src/plugins/mcp/providers/ssh_provider.py
@@ -360,6 +360,21 @@ class SSHProvider(MCPToolProvider):
                     ),
                 )
 
+        # P1 fix 2026-04-11: 群組 B 寫入工具必須有 known_hosts，否則拒絕執行
+        # 防止 MITM — 讀取工具（Group A）允許 known_hosts=None 方便診斷；寫入操作不行
+        import os as _os
+        if tool_name in GROUP_B_TOOLS:
+            _kh = _os.environ.get("SSH_MCP_KNOWN_HOSTS_FILE")
+            if not _kh or not _os.path.exists(_kh):
+                return MCPToolResult(
+                    success=False,
+                    execution_id=execution_id,
+                    error=(
+                        "Group B write tool refused: SSH_MCP_KNOWN_HOSTS_FILE not set or missing. "
+                        "Set up known_hosts per docs/runbooks/ssh-mcp-setup.md before write operations."
+                    ),
+                )
+
         # 執行
         try:
             is_group_b = tool_name in GROUP_B_TOOLS
diff --git a/apps/api/src/services/decision_manager.py b/apps/api/src/services/decision_manager.py
index aaa302ef..a3b9617c 100644
--- a/apps/api/src/services/decision_manager.py
+++ b/apps/api/src/services/decision_manager.py
@@ -41,6 +41,23 @@ logger = structlog.get_logger(__name__)
 # Phase 7.5: Playbook 優先閾值
 PLAYBOOK_SIMILARITY_THRESHOLD = 0.85  # 相似度 >= 85% 直接使用 Playbook
 
+# P1 fix 2026-04-11: background task GC guard — keep strong refs until done
+_background_tasks: set[asyncio.Task] = set()
+
+def _fire_and_forget(coro) -> asyncio.Task:
+    """Create a background task with GC protection via _background_tasks."""
+    task = asyncio.create_task(coro)
+    _background_tasks.add(task)
+    task.add_done_callback(_background_tasks.discard)
+    return task
+
+# P1 fix 2026-04-11: kubectl action dangerous char whitelist
+import re as _re_module
+_ALLOWED_KUBECTL_PATTERN = _re_module.compile(
+    r"^kubectl\s+(rollout restart|rollout undo|scale|delete pod|get|describe|logs)"
+    r"\s+[a-zA-Z0-9_./-]+(\s+(-n|--namespace)\s+[a-zA-Z0-9_-]+)?$"
+)
+
 
 # =============================================================================
 # Phase 31 (ADR-067 2026-04-10): Log 異常摘要 — NemoTron deepseek-r1:14b
@@ -238,13 +255,13 @@ async def _push_decision_to_telegram(
 
         # Phase 31 (ADR-067 2026-04-10): Log 異常摘要 — NemoTron deepseek-r1:14b
         # 非同步執行，不阻塞主流程
-        asyncio.create_task(_send_log_summary(incident))
+        _fire_and_forget(_send_log_summary(incident))
 
         # MCP Phase 4a: NemoClaw second opinion (2026-04-11 Claude Sonnet 4.6)
         # 若 proposal_data 有 advisory_note，用 NemoClaw bot 身分追加一條訊息
         _advisory_note = proposal_data.get("advisory_note", "")
         if _advisory_note:
-            asyncio.create_task(
+            _fire_and_forget(
                 gateway.send_as_nemotron(
                     f"🤔 <b>NemoClaw 第二意見</b> (信心={confidence:.2f})\n"
                     f"<i>{_advisory_note}</i>"
@@ -446,37 +463,29 @@ async def _fetch_metrics_snapshot(incident: Incident) -> dict:
         if alertname in ("HostHighCpuLoad", "HostOutOfMemory"):
             if instance:
                 host = instance.split(":")[0]
-                r = await prom._instant_query(
-                    f'100 - (avg by(instance) (irate(node_cpu_seconds_total{{mode="idle",instance=~"{host}.*"}}[5m])) * 100)'
-                )
-                if r.get("status") == "success":
-                    for item in r.get("data", {}).get("result", []):
-                        snapshots["cpu_pct"] = round(float(item["value"][1]), 1)
-            r2 = await prom._instant_query(
+                # P0 fix 2026-04-11: _instant_query 要求 dict，回傳 {"result": [...]}
+                r = await prom._instant_query({"query": f'100 - (avg by(instance) (irate(node_cpu_seconds_total{{mode="idle",instance=~"{host}.*"}}[5m])) * 100)'})
+                for item in r.get("result", []):
+                    snapshots["cpu_pct"] = round(float(item["value"][1]), 1)
+            cpu_query = (
                 f'(1 - (node_memory_MemAvailable_bytes{{instance=~"{instance}"}} / node_memory_MemTotal_bytes{{instance=~"{instance}"}})) * 100'
                 if instance else "100 * (1 - node_memory_MemAvailable_bytes / node_memory_MemTotal_bytes)"
             )
-            if r2.get("status") == "success":
-                for item in r2.get("data", {}).get("result", []):
-                    snapshots["mem_pct"] = round(float(item["value"][1]), 1)
+            r2 = await prom._instant_query({"query": cpu_query})
+            for item in r2.get("result", []):
+                snapshots["mem_pct"] = round(float(item["value"][1]), 1)
 
         elif alertname == "HostOutOfDiskSpace":
-            r = await prom._instant_query(
-                'max(100 - ((node_filesystem_avail_bytes{fstype!="tmpfs"} / node_filesystem_size_bytes{fstype!="tmpfs"}) * 100))'
-            )
-            if r.get("status") == "success":
-                for item in r.get("data", {}).get("result", []):
-                    snapshots["disk_pct"] = round(float(item["value"][1]), 1)
+            r = await prom._instant_query({"query": 'max(100 - ((node_filesystem_avail_bytes{fstype!="tmpfs"} / node_filesystem_size_bytes{fstype!="tmpfs"}) * 100))'})
+            for item in r.get("result", []):
+                snapshots["disk_pct"] = round(float(item["value"][1]), 1)
 
         elif alertname in ("PodRestartingTooMuch", "PodCrashLoopBackOff"):
             pod = labels.get("pod", labels.get("component", ""))
             if pod:
-                r = await prom._instant_query(
-                    f'sum(kube_pod_container_status_restarts_total{{namespace="awoooi-prod",pod=~"{pod}.*"}})'
-                )
-                if r.get("status") == "success":
-                    for item in r.get("data", {}).get("result", []):
-                        snapshots["restart_count"] = int(float(item["value"][1]))
+                r = await prom._instant_query({"query": f'sum(kube_pod_container_status_restarts_total{{namespace="awoooi-prod",pod=~"{pod}.*"}})'})
+                for item in r.get("result", []):
+                    snapshots["restart_count"] = int(float(item["value"][1]))
 
         return snapshots
     except Exception as _e:
@@ -953,12 +962,12 @@ class DecisionManager:
                 await self._save_token(token)
 
                 # 觸發自動執行 (非阻塞)
-                asyncio.create_task(
+                _fire_and_forget(
                     self._auto_execute(incident, token)
                 )
             else:
                 # 需人工審核: 推送到 Telegram
-                asyncio.create_task(
+                _fire_and_forget(
                     _push_decision_to_telegram(incident, token.proposal_data)
                 )
 
@@ -989,18 +998,20 @@ class DecisionManager:
         # 另外：若 target 等於 alertname，代表 LLM 把告警名稱填入 deployment_name，也拒絕
         _alertname = incident.signals[0].labels.get("alertname", "") if incident.signals else ""
         _target_is_alertname = bool(_alertname and _target == _alertname)
-        if "unknown" in action or _re.search(r"[<{][^>}]+[>}]", action) or _target_is_alertname:
+        # P1 fix 2026-04-11: kubectl action 危險字元白名單 — 防止 && || ; > | 注入
+        _action_safe = bool(_ALLOWED_KUBECTL_PATTERN.match(action.strip()))
+        if "unknown" in action or _re.search(r"[<{][^>}]+[>}]", action) or _target_is_alertname or not _action_safe:
             logger.warning(
                 "auto_execute_blocked_unresolved_placeholder",
                 incident_id=incident.incident_id,
                 action=action,
                 target=_target,
-                reason="action 含未解析的 placeholder、unknown、或 target==alertname，拒絕執行",
+                reason="action 含未解析的 placeholder、unknown、target==alertname、或危險字元，拒絕執行",
             )
             token.state = DecisionState.ERROR
             token.error = f"Auto-execute blocked: unresolved placeholder in action: {action[:80]}"
             await self._save_token(token)
-            asyncio.create_task(
+            _fire_and_forget(
                 _push_auto_repair_result(incident, action, success=False,
                                          error="無法確認 deployment 名稱，請人工確認後手動執行")
             )
@@ -1057,7 +1068,7 @@ class DecisionManager:
             )
 
             # 2026-04-09 Claude Sonnet 4.6: 執行成功 → 發 Telegram 結果通知
-            asyncio.create_task(
+            _fire_and_forget(
                 _push_auto_repair_result(incident, action, success=True)
             )
 
@@ -1072,10 +1083,10 @@ class DecisionManager:
             await self._save_token(token)
 
             # 2026-04-09 Claude Sonnet 4.6: 執行失敗 → 發 Telegram 失敗通知 + fallback 人工
-            asyncio.create_task(
+            _fire_and_forget(
                 _push_auto_repair_result(incident, action, success=False, error=str(e))
             )
-            asyncio.create_task(
+            _fire_and_forget(
                 _push_decision_to_telegram(incident, token.proposal_data)
             )
 
@@ -1151,7 +1162,7 @@ class DecisionManager:
             return playbook_result
 
         # MCP Phase 4c: Playbook 無命中 → 非同步產生 AI 草稿 Playbook (2026-04-11 Claude Sonnet 4.6)
-        asyncio.create_task(_generate_playbook_draft_if_new(incident))
+        _fire_and_forget(_generate_playbook_draft_if_new(incident))
 
         # Expert System 同步執行 (立即可用)
         expert_result = expert_analyze(incident)