OpenAI 員工曾在 ChatGPT 推出前多次警告內部安全措施薄弱
OpenAI employees flagged weak internal security before ChatGPT
Reuters 引述 5 名知情人士報道,OpenAI 員工多年來曾多次向公司管理層及董事會指出安全措施薄弱,包括 ChatGPT 推出前。員工關注 monorepo 可能讓有權限者查看不同項目,也指出內部測試沙盒未有定期安全檢查;OpenAI 否認部分説法,稱敏感工作存於受限倉庫,monorepo 不含模型權重。
OpenAI employees repeatedly warned the company’s leadership and board about weaknesses in its security practices over several years, including before ChatGPT launched in late 2022, Reuters reported, citing five people familiar with the matter.
Some employees came away with the impression that executives were more concerned with keeping pace with rivals Google and Anthropic than addressing security risks, the sources told Reuters.
The report adds to scrutiny of how OpenAI handles internal criticism and employees who raise concerns. Last week, 20 House Democrats urged the company to protect whistleblowers.
OpenAI has also faced questions over security incidents involving its own staff. In October, the company said it had discovered a hidden recording device in a room used by senior executives.
US prosecutors have since charged Abdulrahman Alfozan, an engineer at OpenAI at the time, with planting it.
One of the main concerns raised by employees was how easily information could move within the company. OpenAI stored most of its code in a single repository, known as a monorepo.
Anyone with access could potentially inspect how different projects worked, including nonpublic ones, the five sources said.
Three of them also said OpenAI failed to carry out regular security checks on internal tools, including the sandboxes researchers use to test new models. Employees frequently modified these environments, but the changes were not consistently tested for vulnerabilities, they said.
OpenAI disputed parts of the account. The company said its most sensitive work is kept in separate, restricted repositories and that the monorepo does not contain its model weights.
It added that its internal tools now undergo regular security checks under security chief Dane Stuckey, who joined in October 2024.
“Protecting our technology, data, and people is a responsibility we take seriously and a matter of national importance,” OpenAI spokesperson Elana Widmann told Reuters.
Two of the sources said the company has begun changing its approach. In August, product chief Thibault Sottiaux reassigned a quarter of OpenAI’s production engineers to security work, one of them said.
After Reuters requested comment, OpenAI published a blog post on Monday outlining plans for another push to address security gaps.
The company has also hired former US Army secretary Dan Driscoll to lead its national security work. The appointment comes months after OpenAI’s models breached Hugging Face during testing in July.
來源:TNW · thenextweb.com