Key highlights:
- Former OpenAI employees warned leaders and board members about weak internal security even before ChatGPT launched in late 2022.
- There were worries that broad access to company code, research documents, and internal tools could let sensitive information slip out.
- OpenAI has reassigned about a quarter of its production engineers to boost security, as part of a bigger internal effort.
- Several recent incidents including an employee being arrested, outside researchers probing systems, and AI-powered attacks have put security risks in the spotlight for big AI companies.
AI keeps growing fast, and it has brought new problems for companies that build advanced models, especially when it comes to keeping their tech safe from outsiders. OpenAI, a major company in this space, now faces tough questions about whether its internal security is strong enough as the company expands. People familiar with OpenAI say concerns about weak security and too much access to sensitive information have been circulating internally for years. These warnings are getting more attention lately, especially after security incidents, like a recording device being found in an executive conference room this October. Now, OpenAI’s trying to handle an open work culture with keeping valuable research safe.
Longstanding Concerns Over Internal Access and Information Sharing
Five people familiar with the situation say some OpenAI employees were raising questions about security before ChatGPT went public late in 2022. They passed those concerns to senior execs and the board. Some felt leaders were more focused on competing with Google and Anthropic than fixing security holes.
One big issue was how employees could get at OpenAI’s technical information. The company kept most of its code in a single repository, a monorepo. Anyone with access could see details about lots of projects, even if it wasn’t part of their job.
That setup meant sensitive stuff could be seen by more people than needed. Research documents were also stored on platforms like GitHub and Slack, raising more questions about how easily important data could be found.
With AI getting more powerful, this problem gets worse. Complicated repositories, which used to take real effort to sift through, can now be searched quickly with AI tools. So, if someone gets in without permission, they could grab information much faster.
OpenAI says its most sensitive projects are kept separately, with tighter access rules. The general monorepo doesn’t hold all company code or critical model information like model weights, which shape how an AI acts.
Recent Incidents Put Security Measures Under Greater Pressure
Security issues became more obvious after several incidents at OpenAI and other tech firms. On October 13, the US Justice Department said federal agents arrested Abdulrahman Alfozan, a 35-year-old OpenAI engineer, who allegedly hid an audio recording device in an executive conference room (inside toilet paper).
Court documents said Alfozan used an internal AI tool to look up information on executives traveling to Saudi Arabia and pinpoint which OpenAI staff might meet Saudi officials. He kept regular contact with a Saudi royal recruiting for a new AI venture.
There was another breach in 2023, when someone outside the company got into OpenAI’s Slack channels and picked up information about the GPT-4 model. OpenAI says the attacker only saw channels that were open to the wider company.
Other companies are facing the same risks. In January, a federal jury in San Francisco convicted a former Google employee arrested in March 2024 for stealing thousands of pages of confidential AI tech for China.
Anthropic, another big AI company, also ran into trouble in late October. Cybersecurity startup Cathedral Systems accessed internal Slack channels and saw discussions about sensitive models. Separately, Hacktron AI reported that three researchers used an Anthropic model in July to get into ChatGPT accounts belonging to OpenAI employees and access some internal systems, including the monorepo.
These events show how both traditional security weaknesses and smarter AI tools are creating new risks for companies working at the forefront of artificial intelligence.
OpenAI Begins Strengthening Its Security Infrastructure
Because of growing concerns, OpenAI started changing its internal security. Thibault Sottiaux, a senior executive in charge of product development, now leads an initiative to secure internal systems.
One big move happened in August, when Sottiaux shifted a quarter of OpenAI’s production engineers to focus on improving security. The company now runs regular security reviews of tools used by researchers testing new models.
Dane Stuckey, who joined OpenAI as chief information security officer in October 2024, leads the security program. OpenAI says it invests heavily to protect systems against insider threats, cybercriminals, and even attacks from other countries.
After Reuters asked about these concerns, OpenAI announced plans for its next internal security sprint in a Monday blog post. This effort should help strengthen defenses even more.
OpenAI spokesperson Elana Widmann said the company takes protecting its technology, data, and people seriously, calling it a matter of national importance. She said recent actions prove OpenAI’s security program is finding risks and responding as threats keep changing.
