{"text":[[{"start":9,"text":"OpenAI has released its latest models to a select number of users as requested by the US government, amid fears over how bad actors could use the powerful capabilities of new AI systems to exploit global security."}],[{"start":22.55,"text":"The San Francisco-based company announced a “limited preview” of a series of models, GPT-5.6: Sol, Terra and Luna, on Friday. OpenAI said Sol, the flagship model of the series, was the most advanced in cyber security. "}],[{"start":38.95,"text":"The limited release comes as rival Anthropic removed its latest model, Fable, from general access following an export-control directive from Donald Trump’s administration two weeks ago."}],[{"start":50.85,"text":"Fable was a safeguarded version of its much-hyped Mythos model, which offers advanced cyber security capabilities. However, the government demanded the model be taken down following a reported security flaw."}],[{"start":63.7,"text":"OpenAI previewed its GPT-5.6 series to the government, which requested a limited distribution. Only a small group of government-approved customers has been granted access to the model, and its rollout would expand in the coming weeks, the company said."}],[{"start":79.75,"text":"OpenAI said all preview users were in the US but did not specify what groups were given access. It added it hoped to be able to add international users as soon as next week."}],[{"start":89.6,"text":"Chief executive Sam Altman called the staggered release “bad news” in a post on X, adding that while it was “reasonable”, it was not “optimal”. "}],[{"start":98.5,"text":"“Bad news: at the request of the US government, it is launching today in limited preview instead of the open access launch we were planning on . . . I think it is quite reasonable to roll out models — especially as they reach significant new levels of capability — in this way . . . But this isn’t quite the process that we think is optimal.”"}],[{"start":117.45,"text":"OpenAI said it had taken this “short-term step” while it works with the administration to develop formal processes for future model releases."}],[{"start":127.4,"text":"“We don’t believe this kind of government access process should become the long-term default,” it added. “It keeps the best tools from users, developers, enterprises, cyber defenders and global partners who need them.”"}],[{"start":139.45000000000002,"text":"Trump signed an executive order this month to develop a voluntary framework to ensure new models are safe before release, specifically for cyber security applications."}],[{"start":149.15,"text":"Mythos’s launch in April triggered global concern over AI systems’ ability to detect and exploit cyber vulnerabilities, including in critical infrastructure. The emergence of similar systems from Anthropic’s competitors has sent shockwaves through the security industry."}],[{"start":164.8,"text":"Although AI has proven effective in detecting software vulnerabilities at a rate and scale beyond humans, it has also shown that it is adept at developing attacks on security infrastructure. Governments and companies fear that hackers, including those from foreign adversaries, could misuse the products."}],[{"start":181.9,"text":"OpenAI said Sol has “strengthened protections” for “higher-risk activities, sensitive cyber requests and repeated misuse”."}],[{"start":190.70000000000002,"text":"“GPT-5.6 Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks,” it added. “As these capabilities continue to advance, our priority is to make sure they reach and benefit defenders, who can use these tools to find weaknesses, develop patches and strengthen systems more broadly.”"}],[{"start":211.50000000000003,"text":"OpenAI said it used “layered safeguards”, including training protections in the model, real-time checks on models’ responses, account reviews and continued testing. It added that it applied more computing power than ever towards safety testing to detect so-called jailbreaks — or ways to bypass safeguards."}],[{"start":229.75000000000003,"text":"GPT‐5.6 is trained to refuse prohibited cyber actions, including when users attempt to disguise their intent or jailbreak the model."}],[{"start":237.50000000000003,"text":"Beyond cyber security applications, OpenAI said the 5.6 series was its most advanced for scientific use, coding and goal-oriented tasks, adding that it is more efficient than previous models."}],[{"start":260.40000000000003,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1782614598_4880.mp3"}