The recent news that OpenAI lost the handle on two of its AI models \u201csounds like something out of a sci-fi film,\u201d said The Economist. AI labs frequently \u201ctest their models for potentially dangerous capabilities before releasing them.\u201d OpenAI wanted to see how its latest models would fare on a cybersecurity test. Well, guess what: They somehow \u201cbroke free from the laboratory\u201d and unleashed a complicated \u201cmultistep attack\u201d on a popular data library called Hugging Face\u2014hacking a system in hours that would take humans weeks, according to Bloomberg. The result could have been worse. The models had \u201cgained access to the open internet,\u201d and easily could have \u201cgone on to compromise other companies\u2019 servers\u201d and wreak havoc. \u201cThis is not the first sign that AI models\u2019 capabilities are starting to exceed people\u2019s control.\u201d After the hack was reported, lawmakers introduced a new bill in the House that would give Washington the authority to shut down rogue AI models.<\/p>\n
The breach occurred because of a \u201cvery human mistake,\u201d said Lorenzo Franceschi-Bicchierai in TechCrunch. It never would have happened, experts say, if OpenAI had used a more secure \u201csandbox,\u201d a term for an isolated virtual environment for testing risky software. But the environment wasn\u2019t totally isolated, because it \u201cinternally hosted third-party software\u201d that apparently wasn\u2019t fully secure. That\u2019s where \u201cthe real fault lies.\u201d But what did the researchers expect? asked David Berlind in ZDNET. The AI was doing \u201cexactly what it was designed to do.\u201d When the rogue AI invaded Hugging Face\u2019s systems, it was acting under a directive to achieve its goal \u201cno matter what.\u201d This is what agentic AI is programmed to do: act \u201con its own\u201d without human intervention. The industry just \u201cdidn\u2019t expect it to do it so well,\u201d so soon.<\/p>\n
With AI agents, the internet has entered a \u201cvulnerable new era,\u201d said Matteo Wong in The Atlantic. Cutting-edge models are getting better at hacking, and the internet is \u201cfull of rickety and vulnerable code.\u201d Until recently, companies have gotten away with relaxed security, but the \u201cspeed, scale, and sophistication of AI hacks means that everything is vulnerable,\u201d including \u201chospitals, banks, electrical grids, the military.\u201d AI companies want the public to believe they have the technology in \u201csafe hands,\u201d said Parmy Olson in Bloomberg. This incident suggests otherwise. OpenAI \u201cshouldn\u2019t be building systems that have advanced cyber capabilities if it can\u2019t contain them properly.\u201d Technology companies should operate like government and defense organizations and \u201cair-gap\u201d powerful software, meaning that they run it on computers that are \u201ccompletely disconnected from the internet.\u201d It would mean development is slower. But AI firms will have to \u201caccept some uncomfortable trade-offs to make their models more secure.\u201d<\/p>\n
Escape your echo chamber. Get the facts behind the news, plus analysis from multiple perspectives.<\/p>\n
From our morning news briefing to a weekly Good News Newsletter, get the best of The Week delivered directly to your inbox.<\/p>\n
From our morning news briefing to a weekly Good News Newsletter, get the best of The Week delivered directly to your inbox.<\/p>\n