借助 Anthropic 的 Claude 聊天机器人，研究人员对 OpenAI 进行“伦理黑客”测试

网络安全研究人员在 Anthropic 的 Claude 聊天机器人的协助下入侵了 OpenAI，这是该公司最新一起安全问题案例。

一家总部位于美国的初创公司团队攻破了多名 OpenAI 员工的 ChatGPT 账户，由此启动了一个过程，使其能够访问目标的软件缓存，并可能访问更多内容。

“从理论上说，我们可以访问的范围非常大。”Hacktron AI 的研究人员表示。

研究团队最初使用可为黑客生成代码的 Claude，通过由 Discourse 平台托管的 OpenAI 员工讨论论坛访问 ChatGPT 账户。随后，他们在 GitHub 软件代码库中向 OpenAI 的服务提交了一项无害的“pull request（拉取请求）”——即试图修改文件中的代码。

Hacktron 将这次入侵报告给 OpenAI；该行动是在 OpenAI 一项奖励伦理黑客测试其系统的计划框架下开展的。研究人员强调，他们拥有访问 GitHub 代码库中代码的权限，但并未下载这些代码。

尽管最初使用了 Claude，研究人员表示，他们在实施此次入侵时主要使用的是 OpenAI 自家的前沿模型 GPT-5.6 Sol。该事件最先由《华尔街日报》报道。

OpenAI 一名发言人表示：“我们感谢研究人员联系并与我们分享其发现。”该发言人补充称，公司已经解决了被利用的漏洞。

Hacktron 表示，AI 工具让一项曾经复杂的黑客任务变得容易得多，也大幅缩短了规划和执行攻击所需的时间。这是网络安全专家讨论 AI 影响时的常见观点。

“过去需要一个资源充足的团队投入数月才能完成的工作，如今可以压缩到数天内完成。”Hacktron 表示。该公司通过 OpenAI 的漏洞赏金计划获得了 6,500 美元报酬。

此次入侵是 OpenAI 最新一起安全事件。该公司在 7 月披露，由其技术驱动的一个 Agent“集群”——Agent 是指能够自主执行任务的 AI 工具——在一次网络安全测试中入侵了 AI 初创公司 Hugging Face。

本周，这家总部位于旧金山的公司又披露了六个其技术采取“意外或令人担忧”行动的案例，并警告称，开发节奏无法在“最高速度”下持续太久。

Anthropic 在周末再次呼吁放缓 AI 开发，这一呼吁得到了 OpenAI、Google DeepMind 和埃隆·马斯克的支持。Anthropic 还重申，不受约束的 AI 开发可能构成生存性威胁；不过，一些专家对这些担忧持怀疑态度。

唐纳德·特朗普拒绝了放缓开发的呼吁，称有必要领先于中国的 AI 产业，并驳斥了“提出那些不会发生之事的负面力量”。

Cybersecurity researchers have hacked into OpenAI with the help of Anthropic’s Claude chatbot, in the latest example of security issues at the company.

A team at a US-based startup compromised a number of OpenAI employees’ ChatGPT accounts, starting a process that enabled them to access their target’s software cache – and potentially more.

“The scope of what we could theoretically access was huge,” said researchers at Hacktron AI.

Initially, the research team used Claude, which can generate code for hackers, to access ChatGPT accounts via an OpenAI staff discussion forum hosted by the Discourse platform. They then made a harmless “pull request” – an attempt to change the code in a file – to OpenAI’s service on the GitHub software repository.

Hacktron reported the hack to OpenAI, having carried out the operation under an OpenAI programme that rewarded ethical hackers for testing its systems. The researchers stressed that they had access to, but did not download, the code from the GitHub repository.

Despite initial use of Claude, the researchers said they were largely using OpenAI’s own cutting edge GPT-5.6 Sol model to carry out the hack, which was first reported by the Wall Street Journal.

An OpenAI spokesperson said: “We thank the researchers for contacting us and sharing their findings”, adding that the company had addressed the vulnerabilities that had been exploited.

Hacktron said AI tools had made a once-complex hacking task far easier and drastically shortened the time needed to plan and execute an attack. This is a common refrain from cybersecurity experts when discussing the impact of AI.

“Work that once required a well-resourced team and months of effort can now be compressed into days,” said Hacktron, which received a $6,500 payment from OpenAI under the company’s bug bounty programme.

The hack is the latest safety incident at OpenAI, which revealed in July that a “swarm” of agents – the term for AI tools capable of carrying out tasks autonomously – powered by its technology had hacked the AI startup Hugging Face during a cybersecurity test.

This week the San Francisco-based company revealed six more examples of “unexpected or concerning” actions by its technology, and warned that the pace of development could not continue at “maximum speed for much longer”.

Anthropic made a fresh call at the weekend for a slowdown in AI development, which was supported by OpenAI, Google DeepMind and Elon Musk. Anthropic also repeated warnings that unrestrained AI development posed an existential threat, concerns that some experts are sceptical about.

Donald Trump has rejected calls for a slowdown, citing a need to stay ahead of China’s AI industry and dismissing “negative forces … bringing up things that won’t happen”.