Concerns are growing about the national security implications of artificial intelligence. As frontier AI models grow more capable, their ability to break free from their controls grows with them.

Case in point: the recent OpenAI and Anthropic incidents.

America is having the wrong data center conversation — and its shooting us in the foot

Your power bill is paying for Silicon Valley’s AI obsession

Wiping your own phone is now a federal crime — if you do it at the airport

According to recent disclosures, OpenAI’s most powerful models not only went rogue and hacked the AI open-source hub Hugging Face but also roamed the internet unchecked for four days and targeted a customer of another AI company. Further, the rogue AI used exposed logins to gain access to at least four “publicly available services” as part of a larger effort to breach Hugging Face. If a human had done the same thing, it would be a felony, and they would be held criminally liable. 

OpenAI was testing its models in a “sandbox” — or a container with limited internet for the models to download on-demand software tools that the models were supposed to be unable to escape from. However, the incident showcases that today’s AI models and agents have increasingly powerful capabilities to break out of these controlled containers and act in ways that were not intended by their developers, otherwise known as AI misalignment.

After this was revealed, Anthropic reviewed its internal cybersecurity........

© Washington Examiner