There is an artificial intelligence too dangerous for public release that could threaten, among other tbings, the security of major financial institutions
It sounds like science fiction, but it’s not. And yet there seems to have been little reaction.
Nasa’s Artemis II mission to the moon has held a prominent place in the news in the last month, but the more significant launch of Anthropic’s Project Glasswing often had just a passing mention on the technology pages.
Glasswing is Anthropic’s plan to head off the danger posed by its new Claude Mythos model, which it claims has “already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser”.
Markets appear to have reacted only a little more than mainstream discourse.
There have been some nervous meetings among bankers, central bankers, governments and tech firms. Nvidia’s share price is up around 13% since Anthropic announced Glasswing on April 7, though the causality is hard to determine given the US and Iran agreed a ceasefire on April 8.
Under the plan, Anthropic is releasing Mythos to a select group of firms who will try to quickly patch as many of the newly identified weaknesses as they can. Recent history suggests they won’t have long before similar capabilities proliferate across other AI labs and likely end up more widely available.
The easy answer is to be cynical. Some have suggested Mythos is not as good as claimed and that Anthropic is restricting usage to build marketing hype or to ease constraints on computing resources.
More credible is to take Anthropic at its word.
The claims it makes are dramatic, but they contain statements that are highly unlikely to be false. Possibly, there’s some extra hype but even if that’s the case, Mythos still marks a leap forward for AI. Finding high severity vulnerabilities in every major operating system and browser is surely already several steps beyond what the cynics expected from the tech.
Anthropic’s credibility is somewhat enhanced by the cast of names involved in Glasswing, though who would want to miss out?
More significantly, Anthropic has provided limited examples of Mythos’ power.
Besides, Glasswing is at least not initially a lucrative venture, given Anthropic is putting on $100m of usage credits to aid the cybersecurity efforts.
Running out of control
It raises the question of where AI’s advance will stop.
At the dramatic end of the spectrum, the viral AI 2027 forecast published in April 2025 is looking fairly accurate so far. It foresees capabilities that continue to accelerate, ending in one scenario with the extinction of humanity in 2030.
You don’t have to go that far to come to a gloomy assessment of where AI is heading, however.
In August 2025, GlobalCapital compared the rise of AI to the early days of the Covid pandemic. Looking at the trend, you could see what was coming but it was so unthinkable as to be unbelievable. So far, that trend is developing.
If Mythos-like capabilities become widely available, it will transform the digital world. AI will be a vital part of cybersecurity, given attackers would have a scalable way to find and exploit vulnerabilities.
That has implications for every firm, across capital markets and beyond. It has implications for the balance of power in the world and between the private sector and governments, placing US tech firms in an even more powerful position.
More broadly than just cybersecurity, sustained improvement in AI capabilities would mean the February sell-off in software will be just the start of a reckoning over which business models will become obsolete. The upending of the economy could reach far and wide.
It’s tempting to search for a saving grace to avoid confronting such profound problems.
There are any number of factors that could temporarily check the rate of AI improvement: a sustained rise in energy prices; slower than expected revenue growth for AI labs; disruption to supply chains to give just three examples.
But it is getting more difficult by the week to escape from the conclusion that there is no imminent ceiling for AI’s already great powers.
This point will get the attention it deserves. But it would be better to pay attention now before a major cybersecurity disaster forces everyone to do so.