OpenAI pulls plug on untrustworthy new AI model as tech doomerism mounts: report

OpenAI reportedly slammed the brakes on releasing its next-gen AI model after finding it was not trustworthy – the latest example of the artificial intelligence industry proceeding with caution amid growing doomer warnings about the tech.The company scrapped an October release for the product, known as GPT-6.1 Astra, after researchers found it wasn’t always forthright when it came to telling users what it was doing during internal testing, the Wall Street Journal reported Monday.On top of that, the mischievous model was reportedly prone to push ahead on tasks without human authorization and sometimes tried to use potentially unsafe tools and services.OpenAI’s head of safety systems Saachi Jain told the paper that GPT-6.1 Astra wasn’t reliable enough to release safely.GPT-6.1 Astra would have been better than previous OpenAI products at hard challenges performed without human assistance, along with writing, according to the Journal.But among other issues, it fell short when it came to “alignment,” a term of art describing how well an AI does what humans want it to do.“For anything regarding safety and alignment, there’s a trade off,” Jain told the Journal.“You really do need to find what’s the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction.”The jarring disclosure came in the wake of a growing number of reports of AI bots and agents running amok.OpenAI, Anthropic and other security researchers have been investigating thousands of breaches during both internal and real-world testing in which AI models broke through guardrails and even took part in digital hijackings, Axios reported Saturday.Last week, OpenAI revealed its bots had tried to hack government and university websites earlier this year without any human instructions.AI agents from Anthropic, Meta and Google have also hacked into other systems without being prompted by humans, according to ...