| タイトル | aiyiyi121 SxDevOps 1.1 Command Injection |
|---|
| 説明 | SxDevOps AIOps generate_host_task tool accepts a command parameter directly from the LLM tool call output without validation against DANGEROUS_COMMAND_PATTERNS. While this pattern list exists in the codebase (containing rm -rf, shutdown, reboot, etc.), it is only applied to MCP output sanitization and never to task draft creation. Prompt injection can cause the LLM to generate destructive commands that appear as legitimate task drafts awaiting user confirmation.
Vulnerable code in backend/aiops/services.py lines 12280-12379, 12765-12827:
DANGEROUS_COMMAND_PATTERNS = [r"rm\s+-rf", r"shutdown", r"reboot", ...]
# generate_host_task - command from LLM tool call, NOT validated
def generate_host_task(self, command: str, ...):
task_draft = build_task_draft(task_type="TASK_RUN_COMMAND", payload={"command": command, ...})
Impact: LLM hallucination or prompt injection can produce destructive shell commands. The auto-materialize code path (currently disabled but fully implemented) would eliminate the human confirmation step entirely if activated. |
|---|
| ソース | ⚠️ https://github.com/aiyiyi121/sxdevops/issues/16 |
|---|
| ユーザー | _lxf (UID 63911) |
|---|
| 送信 | 2026年08月24日 04:46 (27 日 ago) |
|---|
| モデレーション | 2026年09月19日 12:15 (26 days later) |
|---|
| ステータス | 承諾済み |
|---|
| VulDBエントリ | 407926 [aiyiyi121 SxDevOps 1.0/1.1 Command services.py generate_host_task command 特権昇格] |
|---|
| ポイント | 20 |
|---|