Artificial intelligence (AI) catastrophisers had good reason to say I-told-you-so because of two back-to-back incidents. In one, an unreleased OpenAI model testing its capabilities hacked into the systems of a company called Hugging Face. And, more recently, there came reports that some chats and projects users made with Anthropic’s Claude had been indexed by Google for its Search. Claude has a feature of sharing with other users but once turned on, Google got access.
Though these were not strictly private conversations, technology had found an automated loophole to make this data universal. OpenAI immediately addressed its issue by stating that it was working along with Hugging Face. Anthropic’s response was that the data shared needed to be posted somewhere that the public could access for such indexing to happen. They did fix the loophole immediately but the explanation was also construed as a cop-out that thrust the responsibility onto the user. The incidents have led to calls for a rethink on AI security. Hugging Face CEO Clement Delangue made a post on X, asking for radical transparency by releasing all traces of the rogue AI agent and a commitment of $100 million from OpenAI for cyber defence. He added, “The first autonomous agent cyberattack is an unprecedented event. It deserves an unprecedented response!”