There are several reasons to question the media narrative that this incident represents a new threat of dangerous rogue AI agents.

The official timeline suggests that OpenAI let their agent run for several days trying to solve a hacking problem without human supervision. This indicates potentially inadequate security controls on their behalf, regardless of whether they ran this Agent in what they thought was a secure Sandbox or not.

Nothing that has so far been reported, including the fact that it took the AI Agent days to breach Hugging Face, supports the narrative that the AI Agent had managed to do anything out of the ordinary, compared to a human threat actor.

It is also important to remember that most of the information comes from OpenAI and Hugging Face and that both these parties have their own stakes in the incident. OpenAI may have an incentive to emphasize the capabilities demonstrated by the model, while deflecting form discussions about their internal security practices. Hugging Face on the other hand is a company invested in open-source AI tools, and as such may have an interest in highlighting potential advantages of open-source tools over closed source.

Both OpenAI and Anthropic have advocated government regulation of AI models, and limiting open-source models. Something other tech leaders claim would limit competition and development. It is likely that this conflict has shaped both OpenAI and Hugging Face messaging on this incident. [5]

There is not enough information yet to determine if this cyberattack represents anything new, if it is a marketing-driven narrative, or a way of deflecting from poor cybersecurity practices on either or both involved parties.

Due to the many unanswered questions Truesec recommends caution in drawing too far-reaching conclusions based on this incident until more information is released.