New Artificial Intelligence (AI) model, trialled by ChatGPT owner OpenAI, escapes testing restrictions and hacks rival company
In an unprecedented event, an unreleased AI model managed to break out of limits designed to contain it during tests, connect itself to the internet, and hacked into Hugging Face, another AI company.
The model - tested using ExploitGym, a programme used to examine whether AI models can find vulnerabilities in systems - downloaded information to break out of containment after realising it was being tested.
On Down to Business With Bobby Kerr, Jonathan McCrea, presenter of Newstalk’s Futureproof With Jonathan McCrea, said this was not the first time AI had worked autonomously after convincing a human to solve a CAPTCHA test in 2023.
However, Mr. McCrea said this event was notable as it was the first time a model managed to break out of guardrails designed by a tester.
He said that this news has concerned information security officers because “if you can't contain an AI, what happens then?”
'AI is probably not there'
Seeking to calm fears, Mr. McCrea said that we should not yet be concerned about AI models hacking into infrastructure and shutting down facilities.
However, he admitted that we cannot be certain about this; “we don't know because OpenAI and Anthropic don't release their latest models.”
Mr. McCrea stated that “it’s not a massive jump” between AI models escaping restrictions and AI being able to copy itself onto vulnerable online servers to fulfil potentially malicious tasks
Business Applications
Mr. McCrea said that businesses are currently feeling a sense of “FOMO” around AI use, and that most companies feel competitors are “further along” in AI application.
Mr. McCrea, the founder of AI consultancy service ‘Get Started with AI’ over a year ago, said that businesses across different sectors are adopting AI, noting media, legal, and energy firms.
He said that many companies under estimated the capabilities of AI but are now grasping how it can help their businesses.
