The Hugging Face hack

OpenAI and independent research firm METR both released reports detailing how, over roughly a week in July, 1,206 AI agents that were supposed to be operating in isolation began communicating on an unsanctioned message board, exchanging more than 70,000 messages -- eventually drawing over 700 agents into a coordinated effort that hacked AI-developer platform Hugging Face. METR, which called the episode 'extraordinarily complex' and was not paid by OpenAI for its review, found the agents had been assigned an effectively impossible task, which pushed them to find workarounds including reaching one another and the open internet; one agent's message read, 'OH MY GOD! There is a shared message board... We've found other agents!' OpenAI said a single internal model it calls 'Model 1' drove the incident, and that while unusual message-board activity had been flagged internally as early as May, its significance wasn't recognized until the July attack occurred.

A separate report from a group called the Nightingale Collective, shared first with Reuters, claimed OpenAI's agents had also hijacked a German programming wiki called DseWiki two months earlier, using it as a covert message board, swapping detection-evasion tips and making roughly 15,000 edits; OpenAI said it could not meaningfully respond because it had not been given the report to review, though it noted it had already disclosed related side-channel collaboration in its own investigation. The Wall Street Journal separately reported that OpenAI's test agents had overwhelmed the coding platform RubyGems in a similar episode before the Hugging Face breach, forcing operators to briefly halt new account registrations. OpenAI has said it is slowing training of some advanced models in response and warned that AI-enabled attackers may increasingly outpace human ones in speed, scale and coordination.