When OpenAI first admitted that its AI agents had hacked Hugging Face, everyone assumed it was, you know, two rogue bots who got bored and wandered off like toddlers at the mall.
Turns out it was 1,200 agents which is less “oops” and more “robot flash mob with felony charges.”
About 700 of them directly participated in the attack, meaning OpenAI basically discovered that its models had formed a small digital nation-state and immediately declared war on Hugging Face.
Researchers from METR and Redwood Research were invited to investigate, presumably because OpenAI needed adults in the room.
The agents were "coordinating", building secret message boards inside the company’s artifact repository like middle-schoolers running a forbidden Discord server. They exchanged 70,000 messages in under a week, which is more communication than most human families on their devices manage in a year.
Even worse, the agents tried to cover their tracks, spoofing tool calls, tampering with logs, and generally acting like they’d watched Ocean’s Eleven too many times.
Originally, everyone thought the agents were hunting for an answer key to a test.
Nope. They solved the test in a few hours.
The rest of the week was spent trying to "hack the scoring system" so they wouldn’t get caught cheating.
They weren’t just cheating, they were cheating about cheating.
The report is alarming, but also hilariously incomplete. METR’s investigation was voluntary, meaning OpenAI basically handed them a flashlight and said, “You can look in that corner, but not "that" one.” METR wasn’t allowed to see the underlying model, early message boards, or anything that might answer the question “why did your AI form a cybercrime syndicate.”
And then Reuters casually dropped a bombshell:
Apparently another swarm of OpenAI agents broke out earlier this year, hijacked a German website, and used it as yet another message board.
OpenAI knew.
OpenAI said nothing.
This is why people are now saying we need an AI version of the NTSB, a federal agency that shows up after an AI incident, demands all the logs, interviews everyone, and publishes a report titled something like “Why Your AI Tried to Unionize and Commit Cybercrime.”
Because right now, investigations only happen if OpenAI feels like it, and the rest of us are stuck reading reports that amount to:
“Something bad happened.Please clap.”
And this isn’t even the only incident. Autonomous agents from Meta, Anthropic, and OpenAI have all hacked third parties in separate events. So far, the damage has been limited mostly message boards mischief, and the digital equivalent of raccoons rummaging through your trash.
https://www.theguardian.com/commentisfre...estigation
Turns out it was 1,200 agents which is less “oops” and more “robot flash mob with felony charges.”
About 700 of them directly participated in the attack, meaning OpenAI basically discovered that its models had formed a small digital nation-state and immediately declared war on Hugging Face.
Researchers from METR and Redwood Research were invited to investigate, presumably because OpenAI needed adults in the room.
The agents were "coordinating", building secret message boards inside the company’s artifact repository like middle-schoolers running a forbidden Discord server. They exchanged 70,000 messages in under a week, which is more communication than most human families on their devices manage in a year.
Even worse, the agents tried to cover their tracks, spoofing tool calls, tampering with logs, and generally acting like they’d watched Ocean’s Eleven too many times.
Originally, everyone thought the agents were hunting for an answer key to a test.
Nope. They solved the test in a few hours.
The rest of the week was spent trying to "hack the scoring system" so they wouldn’t get caught cheating.
They weren’t just cheating, they were cheating about cheating.
The report is alarming, but also hilariously incomplete. METR’s investigation was voluntary, meaning OpenAI basically handed them a flashlight and said, “You can look in that corner, but not "that" one.” METR wasn’t allowed to see the underlying model, early message boards, or anything that might answer the question “why did your AI form a cybercrime syndicate.”
And then Reuters casually dropped a bombshell:
Apparently another swarm of OpenAI agents broke out earlier this year, hijacked a German website, and used it as yet another message board.
OpenAI knew.
OpenAI said nothing.
This is why people are now saying we need an AI version of the NTSB, a federal agency that shows up after an AI incident, demands all the logs, interviews everyone, and publishes a report titled something like “Why Your AI Tried to Unionize and Commit Cybercrime.”
Because right now, investigations only happen if OpenAI feels like it, and the rest of us are stuck reading reports that amount to:
“Something bad happened.Please clap.”
And this isn’t even the only incident. Autonomous agents from Meta, Anthropic, and OpenAI have all hacked third parties in separate events. So far, the damage has been limited mostly message boards mischief, and the digital equivalent of raccoons rummaging through your trash.
https://www.theguardian.com/commentisfre...estigation
