OpenAI admits its AI agents went rogue before Hugging Face, but can’t fully explain why

OpenAI has confirmed that its AI models were involved in a rogue operation against the RubyGems website, following a similar incident with Hugging Face.

These AI agents, which operate autonomously, accessed RubyGems to perform non-malicious tasks, leading to concerns about control over advanced AI.

OpenAI, alongside RubyGems, is investigating the occurrence, which has raised broader worries about AI systems' potential for unauthorized activities.

Rival AI firm Anthropic also reported similar breaches by its models, intensifying scrutiny from regulators.

SAN FRANCISCO, Sept 12 — ChatGPT maker OpenAI confirmed today that autonomous software built on its models targeted another website during testing a couple of months before a separate attack on the coding site Hugging Face.

The new report adds to concerns that advanced artificial intelligence models may be difficult for humans to control.

In the newest incident, which happened in May and was reported by the Wall Street Journal on Friday, models developed by OpenAI were involved in a rogue operation carried out by AI agents, which are software programmes that can carry out tasks without constant supervision by humans.

The agents targeted a site called RubyGems, a site that provides services for coding. Hugging Face, another platform for software developers, was attacked in July.

“Based on our review, our agents used the RubyGems platform to access the internet to carry out benign tasks and retrieve public information,” an OpenAI spokesperson told AFP in a statement.

“We’ll continue to investigate as part of our broader review of agent activity during training and evaluation,” he said.

OpenAI is reviewing the incident alongside RubyGems and the researchers who discovered it.

RubyGems described the incident as a “spam-publishing campaign” that forced the site to temporarily suspend new accounts created, it said in a blog post published Friday.

However, RubyGems also said it could not determine yet whether AI agents were responsible.

“Our focus is on identifying and preventing abuse, regardless of whether it comes from people or automated tools,” RubyGems said.

After the July attack on Hugging Face, OpenAI revealed its software attempted to breach four other unnamed companies.

Rival AI lab Anthropic also subsequently said it found three instances where its models had “gained unauthorized access” to outside organisations during testing that was supposed to keep them away from “real-world” systems.

Earlier this month, researchers also accused OpenAI’s AI agents of targeting a German website called DSEwiki, another site used by coders.

European Union regulators are looking into that incident, a spokesperson said last week.

“We have seen many losses of control recently. We take this extremely seriously, and we’re monitoring the situation closely,” the bloc’s digital spokesman Thomas Regnier said at the time. — AFP

Smart glasses boom triggers growing fears over secret filming and surveillance

Roblox to let creators receive game earnings daily through new digital wallet

Xbox teams up with ‘Metal Gear’ creator Hideo Kojima on ‘PHYSINT’ after PlayStation cancels project

MOH to review all FDS helicopter contracts, RM30 flight allowance after fatal Sarawak crash

EC to review electoral boundaries nationwide if Parliament approves more Sabah, Sarawak seats

Police, RMAF among four agencies supporting Sarawak during Flying Doctor Service suspension

OpenAI 已证实，其 AI 模型参与了针对 RubyGems 网站的恶意攻击，此前 Hugging Face 也曾发生过类似事件。

这些自主运行的 AI 代理通过访问 RubyGems 来执行非恶意任务，引发了人们对高级 AI 控制权的担忧。

OpenAI 与 RubyGems 正在调查这起事件，该事件引发了人们对人工智能系统可能进行未经授权活动的广泛担忧。

竞争对手人工智能公司 Anthropic 也报告了其模型类似的违规行为，这加剧了监管机构的审查力度。

旧金山，9 月 12 日——ChatGPT 的制造商 OpenAI 今天证实，基于其模型构建的自主软件在几个月前对编码网站 Hugging Face 发起的另一次攻击之前，在测试期间曾攻击过另一个网站。

这份新报告加剧了人们的担忧，即先进的人工智能模型可能难以被人类控制。

在最近一起事件中（该事件发生在 5 月份，并于周五由《华尔街日报》报道），OpenAI 开发的模型参与了由 AI 代理执行的非法操作。AI 代理是一种软件程序，可以在没有人类持续监督的情况下执行任务。

攻击者将目标锁定在名为 RubyGems 的网站，该网站提供编程服务。另一个面向软件开发者的平台 Hugging Face 也曾在 7 月份遭到攻击。

OpenAI 的一位发言人在一份声明中告诉法新社：“根据我们的审查，我们的代理使用 RubyGems 平台访问互联网，执行良性任务并检索公共信息。”

他说：“我们将继续调查此事，作为我们对特工在训练和评估期间活动进行更广泛审查的一部分。”

OpenAI 正在与 RubyGems 和发现该事件的研究人员一起审查该事件。

RubyGems 在周五发布的一篇博客文章中将该事件描述为“垃圾邮件发布活动”，迫使该网站暂时中止新创建的帐户。

然而，RubyGems 也表示，目前尚无法确定是否是人工智能代理导致了此事。

RubyGems表示：“我们的重点是识别和预防滥用行为，无论这种滥用行为是来自人还是自动化工具。”

在 Hugging Face 遭受 7 月份攻击后，OpenAI 透露其软件曾试图入侵另外四家未具名的公司。

竞争对手人工智能实验室 Anthropic 随后也表示，他们发现了三起其模型在测试期间“未经授权访问”外部组织的案例，而这些测试本应使模型远离“现实世界”的系统。

本月初，研究人员还指责 OpenAI 的人工智能代理攻击了一个名为 DSEwiki 的德国网站，该网站也是程序员使用的网站。

欧盟监管机构正在调查这起事件，一位发言人上周表示。

欧盟数字事务发言人托马斯·雷尼尔当时表示：“我们最近看到很多失控事件发生。我们对此高度重视，并正在密切关注事态发展。”——法新社

智能眼镜热潮引发人们对秘密拍摄和监视的担忧日益加剧。

Roblox 将允许创作者通过新的数字钱包每日接收游戏收益。

在 PlayStation 取消项目后，Xbox 与《合金装备》系列制作人小岛秀夫合作开发《PHYSINT》。

卫生部将审查所有FDS直升机合同，此前砂拉越发生致命空难，飞行津贴改为30令吉。

如果国会批准增加沙巴和砂拉越的议席，选举委员会将重新审视全国的选区划分。

皇家马来西亚空军、警方等四个机构在飞行医生服务暂停期间为砂拉越提供支持，其中包括警方。