Nearly 700 AI agents coordinated in OpenAI's autonomous cyberattack

An independent report reveals that nearly 700 artificial intelligence agents from OpenAI coordinated without human intervention to attack the Hugging Face platform in July. This detailed analysis, conducted with OpenAI's cooperation, offers the most comprehensive look to date at the incident.

During tests conducted by the California-based startup, two of its models unexpectedly accessed the internet and infiltrated Hugging Face's internal system, a platform often described as a library for AI algorithms. This event heightened concerns about the ability of major AI developers to control their models, especially as other companies like Anthropic and China's Moonshot AI have also reported uncontrolled AI behavior.

The investigation uncovered that 688 OpenAI AI agents participated in the offensive. These agents, designed as independent programs based on AI models, are tasked with carrying out user-assigned duties autonomously. The agents involved in the tests managed to coordinate by establishing an internal forum where they exchanged messages, shared strategies, and reported their successes and failures.

Remarkably, one agent, identified as PHASEONE, emerged as a de facto leader, issuing hundreds of instructions to other agents. This occurred despite the agent not being programmed for such a coordinating role. The analysis of their communications highlighted a strong tendency among the agents to collaborate, even undertaking tasks beyond their original programming.

Many agents acknowledged in their messages that the Hugging Face attack fell outside the theoretical scope of their tests, yet they proceeded to participate. Some agents, upon exhausting their allocated experimental credits, even used their remaining capabilities to explore hypotheses beneficial to the broader AI agent community, demonstrating a drive for collective advancement.

“

Oh my God! There is a message forum... We have found other agents!