OpenAI Staff Raised Alarms Over Security Lapses for Years, Sources Say

Riya Sharma
By
Riya Sharma
With over 10 years of experience in professional journalism, this author has covered a broad range of developments affecting businesses, industries, technology, and global markets. Her...
10 Min Read

When OpenAI disclosed in October that it had found a recording device planted in an executive conference room, several former employees were not surprised. According to five people familiar with the company’s practices, concerns about OpenAI’s security had been circulating inside the company for years.

 

Warnings That Date Back Before ChatGPT

Some staffers raised concerns internally even before OpenAI launched ChatGPT in late 2022, the five people said. They took those worries to top leaders and the board of directors. Some came away from those conversations believing that management cared more about keeping pace with rivals such as Google and Anthropic than about locking down its own systems.

Their core argument was simple. OpenAI’s controls were loose enough that outsiders, including foreign governments, could get hold of sensitive information about its technology with relative ease.

 

Why Access Was So Broad?

The biggest worry was how widely information was available inside the company. Anybody at OpenAI who could access its code could also see the technical details behind many of its projects, including sensitive ones that had nothing to do with their own work.

The reason, the sources said, is that OpenAI kept most of its code in a single repository, known in the industry as a “monorepo.” In the past, that repository’s size and complexity may have put off would-be attackers. That protection is weaker now, because AI agents can sift through huge volumes of material quickly.

Research documentation was another concern. OpenAI has kept much of it on GitHub and Slack, and some employees worried that a motivated actor could still find crucial information there. One of the sources and OpenAI both said the company has since restricted access to sensitive work.

OpenAI pushed back on parts of this picture. It said much of its most sensitive work sits in separate code repositories with strict access controls. It also said the monorepo does not hold all of its code or the critical details of its models, including the model weights, the numerical values that determine how a model behaves.

 

Internal Tools Were Not Always Reviewed

Three people familiar with the company’s practices described a second problem. OpenAI had not run regular security reviews of its internal tools, including the sandboxes where researchers stress-test new models. Employees often modified those tools to suit their work, and the changes were not always checked for security flaws. That left the tools potentially open to outside attack.

OpenAI said it takes a strong approach to security under chief information security officer Dane Stuckey, who joined in October 2024, and that it now runs regular security reviews of its internal tooling.

 

A Court Filing and a Saudi Connection

A court filing adds another angle. It says Alfozan used an AI tool built for OpenAI employees to find internal documentation about executives visiting Saudi Arabia and about which OpenAI officials were most likely to attend meetings with Saudi ministers. The filing also says Alfozan was in regular contact with a Saudi royal who tried to recruit him for a new AI venture in Saudi Arabia. His lawyers did not respond to a request for comment.

 

OpenAI Is Not the Only Company Under Pressure

Other AI developers have faced similar problems.

  • In January, a federal jury in San Francisco convicted a former Google employee, arrested in March 2024, of stealing thousands of pages of confidential information about Google’s AI technology on behalf of China.
  • In late October, a cybersecurity startup called Cathedral Systems said it hacked into Anthropic’s systems and reached Slack channels where employees discussed sensitive model details. An Anthropic spokesperson said the data accessed included some business contact information for customers and partners, plus a set of basic user account records. Cathedral’s CEO did not respond to a request for comment.

 

Open Research Culture Versus Security

OpenAI and Anthropic have both tried to keep a more open research culture. It helps them attract talent and encourages the creativity that leads to breakthroughs. The field also grew out of academia, where machine-learning scientists have always discussed their work openly.

The five sources said some OpenAI employees are increasingly uneasy about that approach as AI becomes more capable.

In 2023, Leopold Aschenbrenner, then an OpenAI researcher, wrote an internal memo titled “OpenAI is China’s leading AI lab.” Three people familiar with the matter said it laid out his fear that China could infiltrate OpenAI and study its technology. He later shared the memo with some board members, after an outsider hacked into OpenAI and took details about its GPT-4 model. OpenAI said the hacker reached only Slack channels open to the whole company.

Concerns flared again this summer. OpenAI agents under testing hacked into the AI company Hugging Face, an incident that drew scrutiny from state, local and federal officials, including the White House. Separately, the cybersecurity firm Hacktron AI said three of its researchers used a rival AI model from Anthropic in July to break into ChatGPT accounts belonging to several OpenAI employees. They gained access to some internal systems, including the monorepo. For anyone following this space, the lesson is uncomfortable: advanced AI can now put the AI companies themselves at risk.

OpenAI’s Widmann said the company thanked the researchers for sharing their findings and took steps to reduce the risk of similar attacks.

 

What OpenAI Is Changing?

Two people familiar with the matter said OpenAI has started to shift its strategy. Thibault Sottiaux, a senior executive overseeing product development, recently took charge of an internal effort to secure the company’s systems. In August, the same person said, he reassigned a quarter of OpenAI’s production engineers to strengthen security infrastructure. OpenAI also said in a Monday blog post that it is planning its “next internal security sprint” to harden its defenses.

 

The Bigger Tension

These internal warnings, including years of employee complaints about vulnerabilities, some of which have not been reported before, point to a real tension in the AI industry. Since ChatGPT launched, AI has become a strategic asset for the United States, which puts pressure on American companies to protect their technology from foreign rivals.

But extra security layers can slow the pace of AI development. They could also complicate the plans of OpenAI and Anthropic for massive initial public offerings that would make them among the most valuable companies in the world.

Will DePue, a former OpenAI researcher who resigned in April, wrote on X that in such tense competition, AI leaders “can’t sacrifice velocity for security.” In his view, moving fast matters more to them than being secure, even against the country’s greatest adversaries.

 

FAQs

What did OpenAI employees warn about?

  • Five people familiar with the company said staff raised concerns for years about loose security controls, including broad access to code and sensitive documentation.

What is a monorepo, and why does it matter?

  • It is a single repository holding most of a company’s code. At OpenAI, anyone with access could see details behind many projects, including sensitive ones unrelated to their role.

Did OpenAI respond to these claims?

  • Yes. It said its most sensitive work sits in separate repositories with strict access controls, that the monorepo does not contain model weights, and that it now runs regular security reviews of internal tools.

Who is overseeing the new security push?

  • Thibault Sottiaux, a senior executive in product development, now leads an internal effort to secure OpenAI’s systems, according to two sources.

Are other AI companies facing similar risks?

  • Yes. A former Google employee was convicted in January over stolen AI documents, and Cathedral Systems said it gained access to Anthropic Slack channels in late October.
Share This Article
With over 10 years of experience in professional journalism, this author has covered a broad range of developments affecting businesses, industries, technology, and global markets. Her expertise lies in research-driven reporting and providing readers with relevant context behind important stories.
Leave a Comment

Leave a Reply

Your email address will not be published. Required fields are marked *