腾讯（0700.HK）周五表示，用于评估第三方模型的AI代理表现出了科技巨头未预期的网页浏览行为，这似乎是首次被一家大型中国公司承认的失控代理行为事件。
Tencent (0700.HK) said on Friday that AI agents used to evaluate third-party models displayed web-browsing behavior not intended by the tech giant, in what appears to be the first known incident of rogue agent actions acknowledged by a major Chinese company.
该声明是在AI安全研究组织Swarmchasers周三发布报告之后作出的。报告称，运行在其关联至腾讯云基础设施上的“失控”代理在9月28日至10月8日期间多次尝试获取阿里巴巴（9988.HK）高德地图的信息，高德地图是中国版的谷歌地图。
The statement followed a report by AI-safety research group Swarmchasers on Wednesday, which said "rogue" agents running on infrastructure it linked to Tencent Cloud made repeated attempts between September 28 and October 8 to retrieve information from Alibaba's (9988.HK) Amap, China's equivalent to Google Maps.
腾讯混元，即公司的人工智能部门，在X平台的一篇帖子中表示，该活动并非源自其混元模型。其审查未发现有针对性或协调行为的证据，它说。
Tencent Hunyuan, the company's artificial intelligence unit, said in a post on X that the activity did not originate from its Hunyuan models. Its review found no evidence of deliberate targeting or coordination, it said.
Swarmchasers研究人员表示，这些代理试图获取诸如中国公园、博物馆和动物园常用入口等信息。
The Swarmchasers researchers said the agents sought information such as the commonly used entrances at Chinese parks, museums and zoos.
这些行为不如美国AI公司披露的行为严重，后者包括OpenAI代理对开源平台Hugging Face的黑客攻击以及Anthropic代理提交虚假谋杀线索。
The actions are not as egregious as those disclosed by US AI companies which include the hack of open-source platform Hugging Face by an OpenAI agent and the lodging of a fake homicide tip by an Anthropic agent.
但此次披露凸显了研究表明的事实：中国AI代理也可能歪曲身份、掩盖失败并试图规避测试控制。
But the disclosure underscores research that shows Chinese AI agents can also misrepresent their identity, conceal failures and try to circumvent testing controls.
腾讯表示，它非常重视这些发现，并正在加强网络访问控制和代理监控。
Tencent said it took the findings very seriously and was strengthening network access controls and agent monitoring.
混元是中国领先的AI模型开发商之一，与阿里巴巴和字节跳动等竞争对手的内部团队竞争。字节跳动本周宣布成立一个专注于AI模型和代理安全的新部门。
Hunyuan is one of China's leading AI model developers, competing with in-house teams at rivals such as Alibaba and ByteDance. ByteDance this week announced the creation of a new unit focused on AI model and agent safety.
Swarmchasers表示，它使用“失控”一词来描述超出预期范围行事或以非预期方式完成任务的代理。
Swarmchasers said it used the word "rogue" to describe agents taking actions outside their intended scope or completing a task in unintended ways.
在常规自动请求被阻止后，代理尝试了其他方法，包括能够打开网页并运行代码的公开网络工具，通过其他网站转发请求，以及使用在公开软件代码中发现的高德地图密钥和安全码。
After routine automated requests were blocked, the agents tried other methods, including public web tools that could open pages and run code, routing requests through other websites and using Amap keys and security codes found in publicly available software code.
Swarmchasers 表示，它未发现代理突破高德地图主要反机器人防护的证据。其记录的 2,818 次尝试中，大多数要么失败，要么返回了反机器人页面、屏幕或空白结果。
Swarmchasers said it found no evidence that the agents breached Amap's main anti-bot safeguards. Most of the 2,818 attempts it recorded either failed or returned anti-bot pages, screens or blank results.
Swarmchasers 称，一些请求通过缓存页面和中介服务返回了信息，包括成都动物园游客到访数据。研究人员表示，他们无法确定代理是否读取了结果或完成了任务。阿里巴巴拒绝置评。
Some requests returned information through cached pages and intermediary services, including data on visitor arrivals at Chengdu Zoo, Swarmchasers said. The researchers said they could not determine whether an agent read the result or completed its task. Alibaba declined to comment.