OpenAI employees warned leaders and the board for years that weak security controls could expose sensitive AI technology, Reuters reported from San Francisco on 4 November. Five people familiar with the matter said some warnings predated ChatGPT’s launch in late 2022, putting the concerns well before the recent criminal case involving an alleged recording device at OpenAI’s offices.
Key points
- Employees took security concerns to OpenAI leaders and its board, according to five people familiar with the matter.
- Some staff concluded management placed greater weight on keeping pace with Google and Anthropic.
- Workers feared outsiders, including foreign governments, could gain access to sensitive technical information.
OpenAI’s shared code repository
Employees told senior leaders and directors about the risks, and some came away believing management was more concerned with competition from Google and Anthropic, the five people told Reuters. They feared that outsiders, including foreign governments, could obtain information about OpenAI’s technology relatively easily. Staff who had access to company code could also inspect technical details of many projects outside their own work.
Most of that code was held in one shared repository, known as a monorepo, News.Az reported. OpenAI says its most sensitive projects sit in separate repositories with strict access controls. It also says the monorepo excludes model weights, the numerical values that determine how a model behaves.
In 2023, former OpenAI researcher Leopold Aschenbrenner wrote an internal memo titled “OpenAI is China’s leading AI lab” that warned of possible Chinese infiltration, three people familiar with the matter told News.Az. He later shared it with some board members after an outside hacker accessed information about GPT-4. OpenAI said the hacker reached only Slack channels available across the company.
Sottiaux shifts engineers to security work
Thibault Sottiaux, a senior executive overseeing product development, has taken charge of an initiative to secure OpenAI’s systems. In August, he reassigned a quarter of the company’s production engineers to security infrastructure, according to one person cited by News.Az.
Three people familiar with OpenAI’s practices said internal tools used by researchers to test models had not routinely undergone security reviews in the past. Researchers frequently changed those tools, and modifications were not always checked for vulnerabilities. OpenAI says it now reviews its internal tools regularly.
The company has also raised its top bug bounty to $172,000. That measure addresses reports from people looking for vulnerabilities; the employees’ earlier warnings also concerned how widely information could be accessed inside the company.
Alfozan case reaches a grand jury
The security questions now sit alongside a criminal case involving Abdulrahman Alfozan, a former OpenAI engineer accused of placing a concealed audio recorder in an executive conference room at its San Francisco offices. A federal grand jury has indicted him over the alleged office bugging.
News.Az reported that, according to a court filing, Alfozan queried an internal AI tool for records of executive trips to Saudi Arabia and for the names of officials who might join talks with Saudi ministers. The allegation concerns access to information inside OpenAI as well as the device in its meeting room.
Elana Widmann, an OpenAI spokesperson, said the company invests substantially in protection against insider threats, cybercriminals and nation-state attacks. “Protecting our technology, data, and people is a responsibility we take seriously and a matter of national importance,” she said.
