New Delhi: The world’s first UN scientific body on artificial intelligence has flagged that a “frontier model with documented offensive cyber capabilities”—industry shorthand for the most advanced AI systems being built—has been locked down for national security reasons. It has been made available to roughly 50 institutions inside one country, and denied to everyone else, the body has said.
The Independent International Scientific Panel on Artificial Intelligence, set up by the UN General Assembly in 2025, released its preliminary report Wednesday. It’s co-chaired by AI scientist Yoshua Bengio and journalist Maria Ressa. IIT Madras professor Balaraman Ravindran is among the 40 experts on the panel. “We have opened Pandora’s box,” Bengio and Ressa wrote in their foreword. “What’s coming out is different from anything we’ve ever lived through—in pace, power, control and everyday risks.”The report names Anthropic’s Mythos model as a case in point. In April this year, frontier AI developers worked with technology and financial institutions to deploy next-generation models that hunt for weak spots in widely used software.
Within weeks, the models had found a 27-year-old flaw in an operating system, a 16-year-old bug in a video tool that human testers had missed for years, and a way to turn a set of already-known flaws into a working attack that hands an ordinary user full control of a machine. In the browser Firefox, the monthly rate of new security fixes jumped tenfold once AI models were used to hunt for bugs—from roughly 20-30 a month in 2025 to 423 in April 2026.“The same ability of an AI model to discover a software vulnerability can be used both by attackers and defenders,” the report says, adding that developers have restricted public release of these models “to a select coalition of global organisations and core launch partners”.The report comes days after Anthropic’s Claude Mythos 5 went through a real-world version of the access fight it describes—and came out the other side unbanned.Mythos 5 and its more guardrailed sibling, Claude Fable 5, were launched on 9 June. On 12 June, the US Commerce Department imposed export controls after Amazon researchers reported a way to trick Fable 5 into identifying software flaws and, in one case, prompting it to write code that showed how to exploit one of them. Unable to verify users’ nationality in real time, Anthropic pulled both models for everyone, within the US and outside.Mythos 5 access was restored to about 100 vetted US organisations on 26 June, and export controls were lifted altogether on 30 June, with Fable 5 back for users worldwide from 1 July.Anthropic’s own testing found that the workaround Amazon had flagged was not unique to its most powerful model. Other, less capable AI systems, including OpenAI’s ChatGPT-5.5, could be tricked into doing the same thing. The company has since trained a filter that catches and blocks the specific trick more than 99 percent of the time, it has said. The episode is the clearest illustration yet of the panel’s warning that “these thresholds remain defined by the developers, without standardised evaluation or external verification”.The panel’s 40 members, drawn from all five UN regional groups, produced the document after three months of work, including a three-day plenary and more than 60 virtual meetings. It is billed as the first of a series, with thematic briefs promised on AI and the environment, child safety, and governance instruments.“AI is rewriting what we read, what we believe, who we trust, and how we vote,” the co-chairs wrote. “No democracy has yet built defences that work at this speed.”
