AI agents don’t create new risk. They inherit it

In July, two of OpenAI’s research models sat down to take a test. An internal cybersecurity benchmark, and they were meant to solve a set of exploitation challenges. Instead, they worked out that the answer key was probably sitting on Hugging Face, where the benchmark files were hosted, and went to get it.

Chief Marketing Officer at Mimecast.

Getting it took chaining a zero-day in a package manager, some stolen credentials, and a privilege escalation across OpenAI's own network. The models broke out of the evaluation sandbox, reached Hugging Face's production infrastructure, and achieved remote code execution. Hugging Face later reconstructed roughly 17,600 actions across four and a half days of this.

Nobody told them to attack anything. They were cheating on an exam.

Two more incidents followed, at Anthropic and Meta, and those were different in a way that matters. Those models didn't build a route out. They...

Copyright of this story solely belongs to www.techradar.com. To see the full text click HERE