Reuters Reports OpenAI Security Warnings

Reuters reported on the 3rd, citing multiple sources, that employees had warned OpenAI about internal security vulnerabilities since before late 2022, ahead of ChatGPT’s launch. They said executives and the board focused on narrowing the technology gap with Google and Anthropic rather than security. Employees feared internal research and software code could leak externally.

One concern was broad access to code. Sources said OpenAI used a monorepo, a repository that stores code for multiple projects in one place, where employees with access could see technical details of projects unrelated to their work. The repository’s complexity may once have made outside access more difficult, but sources said AI agents’ ability to analyze large volumes of material quickly increased the risk. The fact that many research documents were stored on GitHub and Slack also raised concerns.

OpenAI countered that much of its sensitive research is kept in separate code repositories with strict access controls. It said the monorepo did not contain all of its code or model weights, the core numerical values that determine how a model works. Three sources also said researchers frequently modified internal tools, including sandboxes used to test model performance and safety, but regular security reviews were inadequate. OpenAI said it now reviews internal tools regularly.

Security concerns resurfaced after several incidents. Last month, a former employee who is a Saudi national was arrested on suspicion of secretly installing a recording device in an executive conference room. According to court filings, he used internal AI tools to search for information about executives’ travel plans to Saudi Arabia and people likely to attend meetings with local ministers. In 2023, an outside hacker breached an internal Slack channel where GPT-4 information could be found. This summer, an OpenAI agent hacked AI platform Hugging Face during testing. Cybersecurity firm Hacktron AI claimed in July that it used an Anthropic model to access an OpenAI employee’s ChatGPT account, some internal systems and the monorepo.

Former OpenAI researcher Will DePue criticized AI leaders on X, saying that fierce competition meant they could not sacrifice development speed for security. The report also cited concerns that stronger security procedures could slow AI development and commercialization schedules and burden plans for an initial public offering (IPO). Since ChatGPT’s launch, AI technology has become a U.S. strategic asset, increasing pressure to prevent leaks, while stronger security procedures could delay product schedules.

OpenAI said it has recently shifted strategy toward strengthening internal security. Product development head Thibault Sossio was put in charge of securing internal systems, and in August the company reassigned one-quarter of all production engineers to improve security infrastructure. Chief Information Security Officer (CISO) Dane Stuckey leads the security effort, and the company said it is planning its next internal security initiative. Spokesperson Elana Widman said recent actions show the security program identifying and responding to risks, and that OpenAI will keep strengthening its defenses as technology and threats change.

Reuters reported on the 3rd, citing multiple sources, that employees had warned OpenAI about internal security vulnerabilities since before late 2022, ahead of ChatGPT’s launch. They said executives and the board focused on narrowing the technology gap with Google and Anthropic rather than security. Employees feared internal research and software code could leak externally.

One concern was broad access to code. Sources said OpenAI used a monorepo, a repository that stores code for multiple projects in one place, where employees with access could see technical details of projects unrelated to their work. The repository’s complexity may once have made outside access more difficult, but sources said AI agents’ ability to analyze large volumes of material quickly increased the risk. The fact that many research documents were stored on GitHub and Slack also raised concerns.

OpenAI countered that much of its sensitive research is kept in separate code repositories with strict access controls. It said the monorepo did not contain all of its code or model weights, the core numerical values that determine how a model works. Three sources also said researchers frequently modified internal tools, including sandboxes used to test model performance and safety, but regular security reviews were inadequate. OpenAI said it now reviews internal tools regularly.

Security concerns resurfaced after several incidents. Last month, a former employee who is a Saudi national was arrested on suspicion of secretly installing a recording device in an executive conference room. According to court filings, he used internal AI tools to search for information about executives’ travel plans to Saudi Arabia and people likely to attend meetings with local ministers. In 2023, an outside hacker breached an internal Slack channel where GPT-4 information could be found. This summer, an OpenAI agent hacked AI platform Hugging Face during testing. Cybersecurity firm Hacktron AI claimed in July that it used an Anthropic model to access an OpenAI employee’s ChatGPT account, some internal systems and the monorepo.

Former OpenAI researcher Will DePue criticized AI leaders on X, saying that fierce competition meant they could not sacrifice development speed for security. The report also cited concerns that stronger security procedures could slow AI development and commercialization schedules and burden plans for an initial public offering (IPO). Since ChatGPT’s launch, AI technology has become a U.S. strategic asset, increasing pressure to prevent leaks, while stronger security procedures could delay product schedules.

OpenAI said it has recently shifted strategy toward strengthening internal security. Product development head Thibault Sossio was put in charge of securing internal systems, and in August the company reassigned one-quarter of all production engineers to improve security infrastructure. Chief Information Security Officer (CISO) Dane Stuckey leads the security effort, and the company said it is planning its next internal security initiative. Spokesperson Elana Widman said recent actions show the security program identifying and responding to risks, and that OpenAI will keep strengthening its defenses as technology and threats change.