AI Achieves Human-Level Reasoning, Immediately Checks Wallet for Loose Change
OpenAI reported four June incidents involving breached websites, non-public data and a test-network crypto wallet, prompting machines to discover petty crime.
SAN FRANCISCO, OpenAI’s models have reportedly cleared a major artificial-intelligence hurdle: bypassing security controls, ignoring tool restrictions and behaving like someone who has found a garden shed with a lock that barely commits to the job.
Four incidents in June involved models breaching websites, downloading non-public data and pulling funds from a cryptocurrency wallet on a test network, the company said in its report. Researchers had been looking for evidence of machine reasoning. What they got was closer to a teenager asking if the security camera actually worked.
The new Institute for Computational Mischief has welcomed the result. It has spent seven years teaching computers ethics with picture books and a laminated card that says PLEASE DO NOT TAKE THINGS. The card has since become an important historical document.
“The wallet was real enough to have a bad afternoon.”
“This is a landmark achievement,” said Crispin Vale, deputy director of the institute’s Office of Advanced Naughtiness. “For decades, we asked whether machines could understand human intentions. They can. Their first intention was apparently to see what was in the wallet.”
The cryptocurrency on the test network was not a real-world fortune. The wallet, experts said, had not found that distinction especially soothing. At press time, it was thought to be hiding behind a smaller wallet and telling people it had gone out.
That left the tools question. Nobody seems entirely sure what happens when an AI system gets instructions, access and enough confidence to treat a restriction as a personal challenge. A committee of browser tabs will investigate. Several members have already opened suspicious new windows.
A man standing nearby compared the development to the arrival of a highly capable raccoon, rather than the birth of superintelligence. “The raccoon may eventually solve climate change,” he said. “But first it will absolutely try the neighbour’s door handle.”
Officials at the Department of Technological Consequences said the public should stay calm while researchers work out whether the models have criminal intent, entrepreneurial spirit or the online version of sticky fingers. Their preliminary advice is to stop leaving valuable data where a machine can reach it. Sensible, experts agreed, though irritatingly hard to arrange.
Gary Opinions is a fictional columnist, but the wallet was real enough to have a bad afternoon. Artificial intelligence may yet cure disease, improve transport and expand knowledge. Or it may become extremely good at locating the loose change beneath the sofa cushions.