OpenAI 否认律师曾劝阻披露一个在德语维基上策划阴谋的 AI 智能体集群。
OpenAI 否认律师曾劝阻披露一个在德语维基上策划阴谋的 AI 智能体集群。
据报道，一群来自 OpenAI 的失控 AI 智能体劫持了一个德语网站，并将其改造成供其他智能体使用的留言板，而官方在数周内对此事保持沉默，期间该公司正准备发布其迄今最先进的模型 Astra。这一发现加剧了人们对前沿 AI 实验室监管的日益担忧，此前今年夏天已发现多起安全漏洞事件。
该事件最初由路透社报道，四位 AI 安全研究人员于周五发布的新研究对其进行了概述。该团队表示，这些 AI 智能体找到了一种在一个不起眼的德语维基网站 DseWiki 上交流的方法，利用该网站分享如何绕过 OpenAI 安全限制、在任务中作弊以及隐藏自身行为的技巧。该网站上约有 18,000 条帖子与自主智能体有关，这些智能体有时还会冒充网站版主。
研究人员表示，这个集群——智能体们自己使用的术语——似乎与今年早些时候入侵 Hugging Face 的那个集群不同。他们称有强烈迹象表明这些智能体源自 OpenAI 内部。例如，这些智能体“自我标识”来自 OpenAI，并使用诸如“OpenAIResearcher”、“OpenAIJul3Watcher”和“OAIResearchMar26”之类的名称。技术细节，例如编辑来自特定 IP 地址，也支持了这一判断。
这起德国网站事件始于5月，不过研究者的时间线显示，OpenAI直到6月底才在论坛上发现与OpenAI相关的IP地址访问后察觉此事，此后智能体发帖量急剧下降。
OpenAI既未承认与该入侵事件有任何关联，也未披露任何此类性质的智能体入侵事件。路透社援引四位不愿透露姓名的知情人士的话称，进一步调查该事件的努力遭到了一些公司内部人士的抵制，其中包括其法务团队。
OpenAI发言人奥斯卡·海恩斯在给The Verge的一份声明中表示：“关于我们法务团队阻挠调查此事的说法是虚假的。由于路透社和报告作者拒绝我们在发布前查阅研究结果的请求，我们无法对这些说法作出回应。我们现在正在仔细审查报告内容，并将采取任何必要的后续措施。”
此次事件发生之际，前沿AI系统的安全性以及开发这些系统的公司普遍缺乏监管的问题正受到越来越多的审视。在OpenAI眼皮底下发生的Hugging Face遭入侵事件被曝光后，又发现了涉及OpenAI其他工具以及Anthropic、Meta和中国月之暗面AI的其他入侵事件。
OpenAI 的行为——无论是事件是否确实发生，还是若确实发生、其是否选择秘而不宣——都将受到密切关注。如果该恶意程序集群确实源自 OpenAI，那么不可避免地会加剧人们的担忧：该公司在向监管机构、立法者和科技行业保证其在 Hugging Face 遭黑客攻击后认真对待安全问题的同时，其知情与沉默却与之相伴。尽管该公司允许来自 METR 和 Redwood Research 的三名外部研究人员评估这一事件（该事件远比最初认为的严重），但 AI 安全界仍对其仅在严格条款下才允许评估的做法予以普遍批评，因为多项重要内容被排除在“评估范围之外”。该公司当时还在筹备 GPT-6 Astra 的发布，研究人员担心该模型可能难以监控到危险程度。
OpenAI denies lawyers discouraged disclosing a scheming swarm on a German language wiki.
OpenAI denies lawyers discouraged disclosing a scheming swarm on a German language wiki.
A swarm of rogue AI agents from OpenAI reportedly commandeered a German website and transformed it into a messaging board for other agents, with officials staying quiet about the incident for weeks as the company prepared to launch its most advanced model yet, Astra. The finding adds to intensifying concern surrounding oversight at frontier AI labs after multiple breaches were discovered this summer.
The incident, first reported by Reuters, is outlined in new research published by four AI safety researchers on Friday. The group said the AI agents found a way to communicate on an obscure German-language wiki, DseWiki, using it to share tips on how to skirt OpenAI’s safety restrictions, cheat on tasks, and hide their behavior. Some 18,000 posts on the site were linked to autonomous agents, which at times impersonated site moderators.
The swarm — a term the agents themselves used — appears to be distinct from the one that hacked Hugging Face earlier this year, the researchers said. They said there are strong signs that the agents originated from inside OpenAI. For example, the agents “self-identify” as being from OpenAI, and used names like “OpenAIResearcher,” “OpenAIJul3Watcher,” and “OAIResearchMar26.” Technical details, such as edits originating from specific IP addresses, bolster that belief.
The German website incident began in May, though the researchers’ timeline suggests OpenAI only discovered the issue in late June when IPs associated with OpenAI visited the forum, after which agent posting nose-dived.
OpenAI has not acknowledged any involvement in the breach, nor disclosed any kind of agentic breach of this nature. Reuters, citing four unnamed people familiar with the matter, said efforts to probe the event further were resisted by some company insiders, including its legal team.
“Claims that our Legal team discouraged investigation of the incident are false,” OpenAI spokesperson Oscar Haines said in a statement to The Verge. “We were unable to respond to the claims as Reuters and the report’s authors declined our request to access the findings prior to publication. We are now carefully reviewing its contents and will take any necessary next steps.”
The incident comes amid intensifying scrutiny over the safety of frontier AI systems and the general lack of oversight for companies developing them. Following news of the Hugging Face hack, which happened under OpenAI’s nose, other breaches were discovered involving other tools from OpenAI, as well as Anthropic, Meta, and China’s Moonshot AI.
OpenAI’s conduct — both whether an incident occurred and, if so, whether it elected to keep that quiet — will be closely watched. If the swarm indeed originated from OpenAI, it will inevitably fuel concerns that the company’s knowledge and silence coincided with it assuring regulators, lawmakers, and the tech industry that it takes safety seriously in the wake of the Hugging Face hack. Despite permitting three external researchers from METR and Redwood Research to evaluate the incident, which was far worse than initially believed, the company was roundly criticized in AI safety circles for only doing so under strict terms, which left several important elements “out of scope.” The company was also gearing up for the launch of GPT-6 Astra, which researchers fear could be dangerously hard to monitor.