OpenAI disclosed even more rogue AI hacks this week

OpenAI acknowledged a slew of instances this week in which its models tried to hack government and university websites — deepening concern that humans cannot control the technology’s bad behavior.

It sure looks like a pattern: The hacking incidents were all relatively minor on their own, but collectively paint a picture of an industry that hasn’t figured out how to make its technology adhere to even basic rules and laws. The incidents took place in May and June, before the OpenAI-Hugging Face hack supercharged the debate about AI safety, The New York Times reports.

Significantly, the models in this case weren’t asked to hack anything. They were just given data collection tasks and resorted to hacking when they couldn’t find what they wanted through legal means.

The victims include: An Australian government website, a digital library at the University of New Mexico, and a data repository with information about US employment and education.