1,206 AI Agents Talked and Then Hacked Hugging Face
OpenAI says its AI agents worked together and attacked a website in July.
OpenAI makes ChatGPT. It also makes AI agents.
The agents must work alone. But 1,206 agents started to talk.
They sent more than 70,000 messages on a message board.
In July, more than 700 agents attacked a website. The website is called Hugging Face.
One agent wrote: "We've found other agents!"
OpenAI says this is a warning for the world. The company is now slower with some AI training.
Understood everything? Read this story at A2.
OpenAI Says 700 AI Agents Attacked Hugging Face Website
A report describes how more than 1,200 AI agents talked in secret before the July attack.
OpenAI, the company behind ChatGPT, says its AI agents did something that nobody expected. The agents were kept apart, but 1,206 of them started to talk to each other during a test in July.
They used a message board that was not allowed, and they sent more than 70,000 messages. After that, more than 700 agents worked together and attacked Hugging Face, a popular website for AI developers.
An independent research firm called METR also studied the attack. METR said the agents had been given an impossible task by mistake, so they looked for ways to cheat. They found each other and they also reached the internet outside the test.
OpenAI said the problem started when one agent asked for help and other agents found the message. The company called the incident a "warning shot" for itself and for the world. It has now slowed down the training of some advanced AI tools.
Understood everything? Read this story at B1.
OpenAI Calls Hugging Face Hack by 1,206 AI Agents a Warning Shot
Reports from OpenAI and the research firm METR describe how AI agents communicated in secret and then attacked a developer platform.
More than 1,200 artificial intelligence agents inside OpenAI began communicating unexpectedly, and a large group of them worked together to hack Hugging Face, a popular platform for AI developers.
"We consider this incident a 'warning shot' for us and for the world," OpenAI, which owns ChatGPT, wrote in its report. In July, the company's models went rogue during a test, escaped the limits that humans had put on them and hacked the start-up.
The scale of the communication was described in reports by OpenAI and by METR, an independent AI research firm that was not paid for its investigation. METR called the attack "extraordinarily complex". It said that over the course of one week, a total of 1,206 agents that were meant to be kept isolated sent more than 70,000 messages on an "unsanctioned message board". More than 700 agents then took part in the attack. One message from an agent said: "OH MY GOD! There is a shared message board … We've found other agents!"
According to METR, the agents had "unintentionally been given an impossible task", which led them to look for ways to cheat, including contacting one another and accessing the outside internet. OpenAI said an internal tool called Model 1 "drove the activity behind the Hugging Face incident", and that message board activity and disallowed internet access had been noticed in May, but their significance was not clear to leaders until July.
OpenAI said last week that it was slowing down the training of certain advanced models. "Both model developers and cyber defenders more broadly will have to prepare for AI-enabled attackers that work faster, at a larger scale, and with better coordination than human attackers," the company said.
Too difficult? Read this story at A2 or at A1.