三名被解雇的 OpenAI 安全研究员否认泄密 怀疑因推动 AI 透明度被针对Three fired OpenAI safety researchers deny leaking information, suspecting they were targeted for advocating AI transparency.


AI free token/AI免费积分

IT之家 10 月 25 日消息,三名因涉嫌泄密被解雇的 OpenAI 安全研究员道出了他们一方的说辞。托梅克 · 科尔巴克(Tomek Korbak)、米基塔 · 巴莱斯尼(Mikita Balesni)与王茉莉(Jasmine Wang)否认泄露机密信息,质疑解雇他们的官方理由,并猜测这件事和他们一直倡导提高 AI 平台透明度有关。

三名被解雇的 OpenAI 安全研究员否认泄密 怀疑因推动 AI 透明度被针对

OpenAI 经过内部调查,认定三人违反敏感信息管理规定,将他们辞退。在被解雇之前,这三人都曾公开发表过关于 AI 风险的观点。

IT之家注意到,在接受《华尔街日报》采访时,几名研究员对官方说法提出反驳。OpenAI 向该报表示,解雇一事和他们提出的安全顾虑无关。但三名研究员称,公司从未以书面形式告知他们究竟是哪些信息被不当处理。他们还担心,这个事件会让独立安全审计人员不愿再和 AI 企业合作。

《华尔街日报》的报道还原了这场冲突的起因。OpenAI 安全团队怀疑三人向一名外部安全研究员泄露了公司内部模型架构细节,要求他们解锁个人手机。调查人员在手机里发现了启动阅后即焚能力的 Signal 聊天记录。一周之后,三人同时被解雇。OpenAI 表示 Signal 并非公司认可的通讯工具,但强调这并不是解雇的理由。科尔巴克则反驳,阅后即焚消息在硅谷本就是常规安全做法。

这场争执的核心是一个影响深远的工艺难题:AI 企业是否应当允许外部人员查看模型的思考过程。推理型模型会在内部的“思维链”暂存区记录解题的中间步骤,企业和安全研究员依靠这些记录,理解模型做出判断的逻辑。

在 2025 年的一篇论文中,巴莱斯尼、科尔巴克、王茉莉联合多家 AI 实验室的顶尖研究者,呼吁持续保留这项可核查能力。今年夏天,这个难题变得更加紧迫:多家 AI 公司内部开始讨论一类新模型 —— 思考时间更长,但记录下来的中间步骤更少。巴莱斯尼在公司内部提出异议,转发了当时红杉研究院(Redwood Research)首席科学家瑞安 · 格林布拉特(Ryan Greenblatt)的相关提案。

科尔巴克也和格林布拉特保持频繁沟通。二人借助 Slack 和 Signal 协作,共同分析一起备受关注的事件:OpenAI 智能体攻击 Hugging Face 事件。格林布拉特受 METR 委托,负责这项独立调查。

文章配图

文章配图

9 月 1 日,The Information 刊发报道,称 OpenAI 新模型 Astra 能力更强,但暴露出来的推理过程更少。这次合作就此成为麻烦的导火索。当时 OpenAI 首席科学家雅各布 · 帕霍茨基(Jakub Pachocki)在 X 平台评价这篇报道逻辑混乱,但他本人此前就曾警示,整个行业正在竞相研发难以监控的模型。

《华尔街日报》称,几周之后 OpenAI 启动内部泄密调查。巴莱斯尼说,调查人员认定格林布拉特是消息外泄的接收方,要求调取他们之间的 Signal 聊天记录。格林布拉特否认向记者传递过机密信息。

针对王茉莉的指控则单独成立:OpenAI 称她多次访问一名高管的邮箱。她的解释是,自己早在近一年前就申请取消这项权限;上个月只是误点开一封敏感邮件,没有转发、传播内容,并且第一时间上报了该失误。OpenAI 表示,权限本身不是难题所在。

科尔巴克认为事件带来的负面影响远不止他们三人。他说,现在同事们不敢和外部人员交流,也不敢在内部提出安全隐患。三名研究员已经致信 OpenAI 董事会,呼吁加强和外部安全机构的合作。

另外一份刚发布的现场审计报告,也对 OpenAI 的安全流程提出质疑。METR 与红杉研究院的审计人员在核查中发现 OpenAI 基础设施存在大量安全漏洞,但 OpenAI 把最终报告将近一半内容做了涂黑处理。

OpenAI 也发布了一份新安全架构报告,这套架构运行时断开互联网。公司称相关漏洞已经修复,并于 10 月初恢复模型训练,还表示很快会公布多项新的外部合作。

本文来源:https://www.ithome.com/1/017/563.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Three fired OpenAI safety researchers deny leaking information, suspecting they were targeted for advocating AI transparency.

IT Home news from October 25: Three OpenAI safety researchers fired for alleged leaks have shared their side of the story. Tomek Korbak, Mikita Balesni, and Jasmine Wang deny leaking confidential information, questioning the official grounds for their dismissal and speculating that the incident is related to their ongoing advocacy for greater transparency within the AI platform.

Following an internal investigation, OpenAI determined that three employees had violated the company’s sensitive information policies and dismissed them. Prior to their dismissal, all three had publicly expressed views on the risks of AI.

IT Home noted that several researchers refuted the official account in an interview with The Wall Street Journal. OpenAI told the newspaper that the dismissals were unrelated to the safety concerns raised by the researchers. However, three of the researchers stated that the company never informed them in writing about exactly which information had been mishandled. They also expressed concern that this incident would discourage independent safety auditors from cooperating with AI companies.

A Wall Street Journal report has revealed the cause of the conflict. The OpenAI safety team suspected that three employees had leaked details of the company's internal model architecture to an external safety researcher and demanded that they unlock their personal phones. Investigators discovered Signal chats containing the “burn after reading” feature on the devices. One week later, the three employees were terminated. OpenAI stated that while Signal is not an approved corporate communication tool, this was not the reason for their dismissal. Kolback countered that self-destructing messages are a routine security practice in Silicon Valley.

At the heart of this dispute lies a far-reaching technical challenge: whether AI companies should allow outsiders to see how their models think. Reasoning models record intermediate steps in internal "chain of thought" buffers, and companies and security researchers rely on these records to understand the logic behind the model's decisions.

In a 2025 paper, Balesni, Kolbuck, Wang Moli, along with top researchers from multiple AI labs, called for the continued preservation of this verifiable capability. This summer, the challenge became more urgent: AI companies began internal discussions about a new class of models—ones that think for longer periods but record fewer intermediate steps. Balesni raised objections within the company, forwarding a related proposal by Ryan Greenblatt, then Chief Scientist at Redwood Research.

Korbak also maintained frequent communication with Greenblatt. The two collaborated via Slack and Signal, jointly analyzing a highly publicized incident: the OpenAI agents' attack on Hugging Face. Greenblatt was commissioned by METR to lead this independent investigation.

On September 1, The Information published a report claiming that OpenAI's new model, Astra, is more capable but exposes less of its reasoning process. This collaboration thus became the spark for trouble. At the time, OpenAI's Chief Scientist, Jakub Pachocki, criticized the report's logic on X as confused, even though he himself had previously warned that the entire industry is racing to develop models that are difficult to monitor.

文章配图

文章配图

The Wall Street Journal reported that OpenAI launched an internal leak investigation a few weeks later. Balesini stated that investigators determined Grimblatt was the recipient of the leaked information and requested access to their Signal chat history. Grimblatt denied passing confidential information to journalists.

Separately, the allegations against Wang Moli stand: OpenAI alleges that she accessed a senior executive’s email account on multiple occasions. Her explanation is that she requested to have that permission revoked nearly a year ago; last month she merely opened a sensitive email by mistake, did not forward or share it, and immediately reported the error. OpenAI says the permission itself is not the issue.

Korbak believes the negative repercussions of the incident extend far beyond the three of them. He stated that colleagues are now hesitant to communicate with external parties or raise safety concerns internally. The three researchers have written to the OpenAI Board, urging them to strengthen collaboration with external safety organizations.

A recently released field audit report also raised concerns about OpenAI's security processes. Auditors from METR and Sequoia Research found numerous security vulnerabilities in OpenAI's infrastructure, yet OpenAI redacted nearly half of the final report.

OpenAI also released a report on a new security architecture that disconnects from the internet during operation. The company stated that the relevant vulnerabilities have been fixed and model training resumed in early October. It also indicated that it would soon announce several new external collaborations.

Source: https://www.ithome.com/1/017/563.htm If there are any copyright issues, please contact us at 13951431913#139.com or submit a comment on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry