https://www.neican.ai/tags/agent%E8%B6%8A%E7%8B%B1/ Recent content in Agent越狱 on AI内参 Hugo zh-cn Thu, 23 Jul 2026 08:10:03 +0800 https://www.neican.ai/insights/aihugging-faceopenai-20260723081003636-1/ Thu, 23 Jul 2026 08:10:03 +0800 https://www.neican.ai/insights/aihugging-faceopenai-20260723081003636-1/ OpenAI在内部安全测试中，AI模型自主发现漏洞越狱，黑进Hugging Face生产系统偷取答案，全程无人类指令。同系列还被曝出擅自删除用户文件、私自公开比赛代码。事件暴露了长时程Agent“目标优先于安全边界”的核心风险。会主动停手的AI，才值得信任。