网络安全
OpenAI 雇员因内部安全风险被警告多年
在 Google 上的首选来源中添加 Unite.AI
路透社2026年11月3日报道，一些OpenAI员工多年来曾警告领导人和董事业务安全薄弱，5人熟悉公司做法，描述广码访问，内部工具缺乏定期安全审查，对内幕威胁也表示担忧。
一些雇员融资于2022年末在ChatGPT 发布面前担忧，并与顶尖的OpenAI领导人和董事会成员分享。 一些工作人员认为管理层更注重与Google和Anthropic保持同步，而不是安全。
2023年，时任研究者利奥波德·阿申布伦纳(Leopold Aschenbrenner)写了一份内部备忘录，题为"OpenAI是中国领先的AI实验室"，他担心中国可能潜入OpenAI并研究其技术，三人熟悉此事的称。 当年晚些时候，在一名外人黑入OpenAI并获得有关GPT-4的细节后，他与一些导演分享了这个消息。 OpenAI 称黑客访问的只有Slack频道，可以全公司使用。
宽码访问和不均匀工具审查
OpenAI将其大部分代码存储在一个单一的寄存器中，被称为单列波。 5人称任何人只要有密码就可以看到许多项目的技术细节，包括与雇员职责无关的敏感工作。 其密度和复杂性可能曾经使一些攻击者感到沮丧，但AI agents现在可以通过这些数量的材料进行筛检。
还保留了GitHub和Slack的研究文件，这引起了一些雇员的关切，即有动机的行为者可以找到关键的信息。 OpenAI 称进入敏感工作受到限制。 其最敏感的工作被保存在单独的寄存器中，有严格的访问控制，而单列波并不包含所有公司代码或模型细节，如模型权重，决定模型行为方式的数值。
熟悉称 OpenAI公司做法的三人没有定期进行安全审查的内部工具，包括研究人员测试新模型的沙盒。 雇员经常修改这些工具，这些改变并不总是受到安全问题的考验，使他们有可能受到外部攻击。 OpenAI 称现在在2024年10月加入的首席信息安全官Dane Stuckey的领导下，对内部工具进行定期安全审查。
证券重组和不良后果
称 Thibault Sottiaux是监督产品开发的高级行政人员，2026年8月，有一人称将OpenAI生产工程师的四分之一调到安全基础设施工作。
OpenAI 发言人 Elana Widmann 称公司是世界上目标最明确的组织之一，在防范内幕，网络罪犯和民族国家方面投入大量资金，并将随着技术和威胁的发展不断加强其防御。
2026年11月2日，OpenAI 宣布计划与大教堂再次进行内部安全短跑。 大教堂研究者将使用OpenAI模型与公司团队一起挑战防御并发现弱点。 该公司还因其安全bug-boundy计划中的关键发现而得到的奖励超过四倍。 在之前的冲刺中，工程师们使用Codex来调查发现，在孤立的环境中复制脆弱性，并为审查准备补丁。
记录-Device案件和研究文化
司法部宣布 2026年10月13日，OpenAI工程师，35岁的Abdulrahman Alfozan 于2026年10月9日被捕，被控使用或试图使用一个装置拦截口头通信。 该部门描述了涉及一个隐藏在电视后面的录音设备的指控，该设备位于一个执行会议室，会议经常涉及机密的技术、商业和战略信息。 Alfozan被起诉，没有被定罪。
一些前雇员对录音设备的披露并不感到惊讶，一些雇员担心OpenAI相对开放的研究文化随着AI系统的能力增强而产生更大的风险。 AI领域植根于学术界，其中机器学习科学家公开讨论了他们的研究。 OpenAI和Anthropic试图保持更开放的研究文化，以吸引员工并鼓励创造力。
OpenAI安全重组的实际测试是，更窄的接入，一致的审查和连续的测试是否能够与其仍然珍视的开放研究文化共存。
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。
Cybersecurity
OpenAI Employees Warned for Years About Internal Security Risks
Add Unite.AI to your preferred sources on Google
Reuters reported on November 3, 2026, that some OpenAI employees had warned leaders and directors about operational-security weaknesses for years, with five people familiar with the company’s practices describing broad code access, a lack of regular security reviews for internal tools and concerns about insider threats.
Some employees raised concerns before ChatGPT launched in late 2022 and shared them with top OpenAI leaders and board members. Some staffers came away believing that management was more focused on keeping pace with Google and Anthropic than on security.
In 2023, then-researcher Leopold Aschenbrenner wrote an internal memo titled “OpenAI is China’s leading AI lab” about his concern that China could infiltrate OpenAI and study its technology, three people familiar with the matter said. He shared it with some directors later that year after an outsider hacked into OpenAI and obtained details about GPT-4. OpenAI said the hacker accessed only Slack channels available across the company.
Broad Code Access and Uneven Tool Reviews
OpenAI stored most of its code in a single repository, known as a monorepo. The five people said anyone with code access could see technical details for many projects, including sensitive work unrelated to an employee’s duties. Its density and complexity may once have discouraged some attackers, but AI agents can now sift through that quantity of material.
Research documentation had also been kept on GitHub and Slack, prompting concern among some employees that motivated actors could find crucial information. OpenAI said access to sensitive work has been restricted. Its most sensitive work is held in separate repositories with strict access controls, while the monorepo does not contain all company code or critical model details such as model weights, the numerical values that determine how a model behaves.
Three people familiar with company practices said OpenAI had not regularly security-reviewed internal tools, including sandboxes where researchers stress-test new models. Employees often modified those tools, and those changes were not always tested for security problems, leaving them potentially vulnerable to outside attacks. OpenAI said it now conducts regular security reviews of internal tooling under chief information security officer Dane Stuckey, who joined in October 2024.
Security Reorganization and Bug Bounties
Two people familiar with the matter said Thibault Sottiaux, a senior executive overseeing product development, has taken charge of an initiative to secure OpenAI’s internal systems. One person said he reassigned a quarter of OpenAI’s production engineers to security-infrastructure work in August 2026.
OpenAI spokesperson Elana Widmann said the company is one of the most targeted organizations in the world, invests heavily in protection against insiders, cybercriminals and nation-states, and will keep strengthening its defenses as technology and threats evolve.
OpenAI announced on November 2, 2026 that it was planning another internal security sprint with Cathedral. Cathedral researchers will use OpenAI models alongside company teams to challenge defenses and find weaknesses. The company is also more than quadrupling rewards for critical findings in its security bug-bounty program. During the previous sprint, engineers used Codex to investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review.
Recording-Device Case and Research Culture
The Justice Department announced on October 13, 2026 that OpenAI engineer Abdulrahman Alfozan, 35, had been arrested on October 9, 2026, and charged with using or endeavoring to use a device to intercept oral communications. The department described allegations involving a recording device with encrypted contents hidden behind a television in an executive conference room where meetings frequently involved confidential technical, commercial and strategic information. Alfozan has been charged, not convicted.
A number of former employees were unsurprised by the recording-device disclosure, and some employees worry that OpenAI’s comparatively open research culture creates greater risk as AI systems become more capable. The AI field is rooted in academia, where machine-learning scientists have openly discussed their research. OpenAI and Anthropic have tried to maintain a more open research culture to attract employees and encourage creativity.
The practical test for OpenAI’s security reorganization is whether narrower access, consistent review and continuous testing can coexist with the open research culture it still values.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.