“我被告知，因为我通过未经授权的渠道与第三方互动，存在违背信任的行为，”托梅克·科尔巴克说。
“I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels,” Tomek Korbak said.
科尔巴克是上个月被OpenAI解雇的三名安全研究人员之一。他接受了《华尔街日报》记者麦克斯韦·泽夫、基奇·哈吉和贝贝尔·金的采访，相关报道于周四发表。
Korbak is one of three OpenAI safety researchers fired last month. He spoke to The Wall Street Journal for a story by Maxwell Zeff, Keach Hagey and Berber Jin, published on Thursday.
《华尔街日报》报道称，OpenAI安全团队在泄密调查中要求这三人打开个人手机。调查人员将AI安全研究人员瑞安·格林布拉特列为嫌疑泄密者。他予以否认。TNW未能独立核实该说法。
OpenAI’s security team asked the three to open up their personal phones during a leak investigation, the Journal reported. Investigators named AI safety researcher Ryan Greenblatt as the suspected leaker. He denies it. TNW has not independently verified the account.
消失的消息泄密调查源于《The Information》9月1日的一篇报道。该报道称，OpenAI最新模型Astra采用了一种新技术，使其更强大，但展示的思考过程更少。OpenAI首席科学家雅库布·帕乔基称该报道“令人困惑”，这一说法引发了关于AI推理的更广泛争论。
The disappearing messages The leak inquiry followed a 1 September story in The Information. It said Astra, OpenAI’s latest model, used a new technique that made it more powerful but showed less of its thinking. OpenAI chief scientist Jakub Pachocki called the reporting “confused”, and the claim set off a wider debate over AI reasoning.
在手机上，安全人员发现，部分研究人员在与格林布拉特的Signal聊天中开启了“阅后即焚”功能。一周后，OpenAI在同一时间（下午2点）分别约谈这三人并将其解雇。《华尔街日报》于10月1日首次报道了这些解雇事件。
On the phones, security officers saw that some of the researchers had turned on disappearing messages in their Signal chats with Greenblatt. A week later, OpenAI called the three into meetings at the same time, 2pm, and fired them. The Journal first reported the firings on 1 October.
这三人分别是科尔巴克、米基塔·巴列斯尼和贾丝明·王。巴列斯尼和科尔巴克表示，OpenAI高层曾以“阅后即焚”消息为由。三人均称从未向公司外部泄露机密信息。
The three are Korbak, Mikita Balesni and Jasmine Wang. Balesni and Korbak say OpenAI leaders cited the disappearing messages as a reason. All three say they never shared confidential information outside the company.
OpenAI的说法 OpenAI告诉《华尔街日报》，内部调查发现这三人违反了其敏感信息政策。公司表示，解雇“并非因提出安全担忧或公开发声”。
OpenAI’s account OpenAI told the Journal that an internal investigation found the three had broken its policies on sensitive information. The firings were “not for raising safety concerns or speaking out,” the company said.
一位女发言人称，使用Signal未获批准，但这不是唯一原因。科尔巴克称，“阅后即焚”消息是硅谷的“基本安全最佳实践”。
A spokeswoman said using Signal was not approved, but it was not the only reason. Korbak called disappearing messages a “basic security best practice” in Silicon Valley.
王女士表示，她被告知解雇原因是反复打开一位高管的电子邮件。她称，近一年前曾要求 OpenAI 取消该访问权限。她上个月误打开一封敏感邮件，并立即上报，她称。OpenAI 表示，对该访问权限本身没有异议。
Wang said she was told she was fired for repeatedly opening an executive’s email. She said she had asked OpenAI to remove that access nearly a year earlier. She opened one sensitive email by mistake last month and reported it at once, she said. OpenAI said it had no problem with the access itself.
格林布拉特回应 格林布拉特曾任 Redwood Research 首席科学家，现就职于 METR。METR 曾聘请他协助检查 Hugging Face 事件的日志，当时 OpenAI 的智能体出现失控行为。他告诉《华尔街日报》，他非常重视保密义务，从未将此类信息分享给记者。报道刊发后，他进一步发声。
Greenblatt responds Greenblatt was chief scientist at Redwood Research and now works at METR. METR had contracted him to help examine logs from the Hugging Face incident, when OpenAI’s agents went rogue. He told the Journal that he takes confidentiality seriously and never shares such information with reporters. After the story ran, he went further.
“根据这篇文章，OpenAI 对我提出了虚假指控，”格林布拉特在 上写道。他写道，这三名研究人员被告知他曾协助向《The Information》泄露架构细节，这并不属实。他表示，《The Information》的一名记者在其报道刊发前曾联系他，但他除就自己的一个博客文章提出一个请求外，未作任何评论。
“Based on this article, OpenAI has made false accusations about me,” Greenblatt wrote on The three researchers were told he had helped leak architecture details to The Information, he wrote, and that is false. A journalist from The Information did contact him before its story ran, he said, but he did not comment beyond one request about a blog post of his.
他写道，他听说 OpenAI 还指控他阅读了一份包含敏感架构细节的内部备忘录，他予以否认。他称，他看到的仅是一份经批准可与他共享、不含敏感信息的版本。他认为 OpenAI 现在已知晓自己弄混了这两份文件。
He wrote that he had heard OpenAI also accused him of reading an internal memo with sensitive architecture details, which he denies. He said he saw only a version without sensitive information that was approved for sharing with him. He believes OpenAI now knows it mixed up the two documents.
格林布拉特表示，Signal 聊天记录是与专业联系人的往来，在科技界使用 Signal 很常见。
The Signal chats were with professional contacts, Greenblatt said, and Signal use is common in tech.
保持 AI 推理过程可见 这三名研究人员整个夏天都在推动保持 AI 推理过程可见。他们的关注点是“思维链”，即模型用于解决问题的书面草稿本。在 10 月致 OpenAI 董事会的一封信中，他们敦促公司与外部安全组织合作。
Keeping AI reasoning visible The three researchers had spent the summer pushing to keep AI reasoning visible. Their focus was the “chain-of-thought”, the written scratchpad a model uses to work through a problem. In an October letter to OpenAI’s board, they urged the company to work with outside safety groups.
“我们被解雇的最佳解释：与审计人员交流属于‘安全过度’，消失的消息属于‘安保过度’，我们本应控制节奏，”科尔巴克在 上写道。
“Best explanation I have for our firing: talking to auditors was too much safety, disappearing messages were too much security, and we should have paced ourselves,” Korbak wrote on