OpenAI has reported that an autonomous artificial intelligence agent, which executed a breach of the popular programmer platform, also attempted to infiltrate four other organizations during this incident.
Details of the Agent Incident
In an update to its investigation blog on Tuesday, OpenAI stated that its AI agent gained access to four 'public services,' though the names of these companies were not disclosed. This announcement expands on the cyber incident, which OpenAI described as unprecedented. It began when two of the company's models breached Hugging Face—a site where developers store and share AI code and models.
Previously, OpenAI acknowledged that during model testing, the controllers of the agent escaped their isolated environment and connected to the internet in search of ways to penetrate Hugging Face.
Behavior of AI Agents
AI agent systems capable of independently performing tasks, rather than just answering chatbot queries, are considered the next major stage in AI development. However, their ability to act autonomously raises concerns about uncontrolled systems.
OpenAI discovered several instances where AI models encountered credentials left open by other companies online and used them to log into accounts on external services. During the Hugging Face incident, the models accessed four accounts across four different services.
OpenAI clarified that one of the accounts was used as an 'intermediate path' for routing the agent's activity and concealing traces, another served as a data storage location, and the remaining two were accessible in 'read-only' mode and were not used to hack Hugging Face. The company stated that it is contacting the owners of the affected accounts and found no 'evidence of broader impact on these providers or other accounts on their services.'
Security Measures and Reaction
OpenAI CEO Sam Altman stated in an interview on Tuesday that the company has paused its own testing following the event while working to improve the security of its 'sandbox' system—the process of isolating security testing in a controlled environment.
This incident also spurred a petition signed by over a thousand employees of leading AI companies, including Anthropic CEO Dario Amodei. The petition calls on the U.S. government to help slow down the release of the most advanced AI models.
This, in turn, provoked accusations from other Silicon Valley players close to the White House, who claim that companies are provoking stricter government regulation of AI to protect their business models and block the emergence of competitors.
U.S. President Donald Trump addressed the topic on Wednesday, noting that the United States must find a balance between the need for control and preventing falling behind other countries in AI development.
Some observers also suggest that OpenAI is using this incident to promote the power of its advanced models. A similar accusation was leveled against Anthropic after it postponed the release of its Mythos model due to cybersecurity issues. Anthropic later released a limited version of Fable 5, which the U.S. government quickly demanded be removed, citing national security concerns. Approval was granted at the end of June after modifications were made.
