The cyber-abilities of the latest, most powerful AI models, called “frontier” models in AI jargon, can be a boon for cyber-defenders, but the same abilities can also be used for attack.
Spotting weaknesses in systems, for example, is useful for defenders looking to block hackers, but also for attackers hunting for ways to infiltrate an organisation.
Anthropic has tightly restricted access to its systems on the grounds that they are too powerful to fall into the wrong hands, but OpenAI has allowed some firms in Europe, as well as UK banks, to use its most advanced cyber models.
The firm made the announcement alongside the UN General Assembly which is currently underway in New York.
AI is firmly on the agenda there, and US tech bosses including OpenAI’s Sam Altman and Anthropic’s Dario Amodei are due to speak at the high-level gathering.
United Nations Secretary General António Guterres called for the tech to be regulated in his final address, while US President Donald Trump says it should be renamed “super intelligence”.
Evidence suggests the technology is increasingly being used in the Russia-Ukraine conflict.
In a report published earlier this month, Anthropic claimed that its Claude platform appeared to have been used by a group of Russian developers to build software for a swarm of so-called “kamikaze” drones designed for attacks in Ukraine.
It said the group had activated virtual private networks – VPNs – in order to get around a geographical block put in place by the tech firm.
Anthropic also claimed AI had been used in attempts to infiltrate the Ukrainian government, military and diplomatic services.
Recently, around 100 US companies signed an open letter warning governments and institutions around the world that “the window is closing” to ensure cyber defences are robust enough to withstand AI-enabled cyber attacks.
But it’s not just bad actors deploying the tech that are of growing concern: in the summer, OpenAI revealed that a group of AI agents it had been testing had escaped from their controls, and secretly worked together to hack to another tech firm, Hugging Face.
Anthropic, and Google subsequently revealed similar incidents.
Additional reporting by Chris Vallance
