三名OpenAI安全研究者说，他们从未被告知他们实际泄露了什么
关键点
- 被开除的三名OpenAI安全研究人员拒绝分享机密信息。
- 他们怀疑他们的解雇与他们使AI模型更容易监测的努力有关，并警告说，该案可能会阻止雇员提出安全关切。
- OpenAI列举了据称违反政策的行为。 研究者说，调查人员专注于他们与外部研究人员的信号聊天，尽管OpenAI 称信号并不是射击的原因。
三名OpenAI安全研究者因被指控的泄露而被解雇，他们正在讲述他们的故事。 Tomek Korbak,Mikita Balesni，和Jasmine Wang否认分享机密信息，质疑终止他们关系的官方推理，并怀疑与他们主张更透明的AI系统有关。
OpenAI在对其处理敏感信息的政策遭到违反的指控进行内部调查后，将这三人解职。 他们每个人都公开谈论大赦国际的风险，然后才放手。
在接受华尔街日报的采访时，研究人员现在对该说法表示异议。 OpenAI告诉《日刊》说，射击与研究人员对融资的安全关切无关。 但三个人说他们从未被书面告知哪些信息他们应该处理不当。 他们还担心这种情况会妨碍独立的安全审计员完全与AI公司合作。
《日刊》的报道揭示了导致对抗的原因。 OpenAI的安保团队怀疑3人与外部安全研究员分享公司内部模型架构的细节，要求他们解锁个人手机。 在这些设备上，小组发现信号聊天已启用，信息已消失。 一周后，三人同时被解雇。 OpenAI 称 Signal不是一个经过批准的通信工具，但坚持认为这不是终止的原因。 Korbak反驳道，信息消失是整个硅谷的例行安全做法。
关于监督大赦国际推理步骤的争议
争端集中于一个具有重大后果的技术问题。 AI公司是否应该让外人检查他们的模型如何思考？理性模型可以在被称为思维链的内部抓图上记录其解决问题过程的中间步骤。 公司和安全研究者依靠这些痕迹来了解模型是如何做出决策的。
在2025年的一篇论文中，巴勒斯尼，科尔巴克，王与几个AI实验室的主要研究人员一起呼吁继续获得这种能力。 这个问题在夏天变得更加紧迫，当时AI公司的员工开始讨论那些思考时间更长但记下来中间步骤较少的模型。 巴列斯尼内部推后，分享了时任红杉研究首席科学家瑞安·格林布拉特的相关提议。
科尔巴克也经常与格林布拉特接触。 两人在Slack和Signal上空盘旋，同时共同分析一起引人注目的事件——OpenAI agents对Hugging Face的攻击——为此，Greenblatt被METR公司雇用进行独立调查。
关于Astra调查的新闻报道
在9月1日"信息报"报道OpenAI的新型号Astra能力较强但揭示的推理过程较少后，这种合作成为了负债。 OpenAI首席科学家贾库布·帕乔基(Jakub Pachocki)认为报告在当时的X上是泥潭，尽管他本人曾警告过整个行业对不再能够监测的模型的竞争。
几周后，据"日刊"报道，OpenAI 发布是一次内部泄密调查。 巴列斯尼称调查人员将格林布拉特命名为疑似来源，并要求与他见识信号聊天。 格林布拉特否认向记者传递机密信息。
Wang面临单独指控：OpenAI 称她多次访问一个执行官的收件箱。 称在近一年前要求取消访问，上个月意外打开敏感邮件，OpenAI 称访问本身不是问题所在。
根据科尔巴克的说法，沉降波远远超出了它们三个。 同事们现在害怕和外人谈话或提出内部的安全顾虑，他称。 三位研究人员在给OpenAI董事会的信中呼吁与外部安全组织进行更密切的合作。
另外，最近公布的现场审计融资本身对OpenAI安全做法的质疑。 METR和Redwood Research的工作人员在审查期间发现OpenAI的基础设施存在许多安全漏洞，但OpenAI编辑了近一半的最后报告。
OpenAI还发布了自己关于一个新的安全架构的报告，该架构没有互联网接入，称则固定了问题，并于10月初恢复培训。 称公司将很快宣布若干新的外部伙伴关系。
AI News Without the Hype – 由人类治愈
订阅THE DECODER免费阅读，每周AI通讯，我们独家的"AI雷达"前沿报告每年6次，全面存档访问，访问我们的评论部分。
现在订阅
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Three fired OpenAI safety researchers say they were never told what they actually leaked
Key Points
- Three OpenAI safety researchers who were fired deny sharing confidential information.
- They suspect their terminations were tied to their efforts to make AI models easier to monitor and warn that the case could discourage employees from raising safety concerns.
- OpenAI cited alleged policy violations. The researchers say investigators focused on their Signal chats with outside researchers, though OpenAI said Signal was not the reason for the firings.
The three OpenAI safety researchers who were fired over alleged leaks are telling their side of the story. Tomek Korbak, Mikita Balesni, and Jasmine Wang deny sharing confidential information, question the official reasoning behind their termination, and suspect a connection to their advocacy for more transparent AI systems.
OpenAI dismissed all three after an internal investigation into alleged violations of its policies on handling sensitive information. Each of them had spoken publicly about AI risks before being let go.
In interviews with the Wall Street Journal, the researchers now dispute that account. OpenAI told the Journal that the firings had nothing to do with safety concerns the researchers had raised. But the three say they were never told in writing which information they supposedly mishandled. They also worry the case could discourage independent safety auditors from working with AI companies altogether.
The Journal's reporting sheds light on what led to the confrontation. OpenAI's security team suspected the three of sharing details about the company's internal model architecture with an external safety researcher and asked them to unlock their personal phones. On those devices, the team found Signal chats with disappearing messages enabled. A week later, all three were fired at the same time. OpenAI said Signal was not an approved communication tool but insisted it was not the reason for the terminations. Korbak counters that disappearing messages are a routine security practice across Silicon Valley.
A dispute over monitoring AI reasoning steps
The dispute centers on a technical question with major consequences. Should AI companies let outsiders inspect how their models think? Reasoning models can record the intermediate steps of their problem-solving process on an internal scratchpad known as the chain of thought. Companies and safety researchers rely on these traces to understand how models arrive at their decisions.
In a 2025 paper, Balesni, Korbak, and Wang joined leading researchers from several AI labs in calling for continued access to this capability. The issue grew more urgent over the summer, when employees at AI companies began discussing models that think longer but write down fewer intermediate steps. Balesni pushed back internally by sharing a related proposal from Ryan Greenblatt, then chief scientist at Redwood Research.
Korbak, too, was in regular contact with Greenblatt. The two communicated over Slack and Signal while working together on the analysis of a high-profile incident—the attack by OpenAI agents on Hugging Face—for which Greenblatt had been hired by METR to conduct an independent investigation.
A news report about Astra kicked off the investigation
That cooperation became a liability after The Information reported on September 1 that OpenAI's new model Astra was more capable but revealed less of its reasoning process. OpenAI chief scientist Jakub Pachocki dismissed the report as muddled at the time on X, even though he himself had warned of an industry-wide race toward models that can no longer be monitored.
A few weeks later, OpenAI launched an internal leak investigation, according to the Journal. Balesni says investigators named Greenblatt as the suspected source and demanded to see the Signal chats with him. Greenblatt denies passing confidential information to journalists.
Wang faced a separate accusation: OpenAI said she had repeatedly accessed an executive's inbox. She says she had asked nearly a year earlier for that access to be revoked, accidentally opened a sensitive email the previous month, did not share its contents, and reported the incident right away. OpenAI says the access itself was not the issue.
The fallout extends well beyond the three of them, according to Korbak. Colleagues are now afraid to talk to outsiders or raise safety concerns internally, he says. In a letter to OpenAI's board, the three researchers call for closer cooperation with external safety organizations.
Separately, a recently published on-site audit raised its own questions about OpenAI's safety practices. Staff from METR and Redwood Research found numerous security gaps in OpenAI's infrastructure during the review, but OpenAI redacted nearly half the final report.
OpenAI also published its own report on a new security architecture that operates without internet access and said it fixed the issues and resumed training in early October. The company says it will announce several new external partnerships soon.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
Auto-fetched and translated by AI Atlas. All rights belong to the original author.