A powerful new Artificial Intelligence model developed by Anthropic has triggered crisis discussions among finance ministers, central bankers and senior financiers over fears it could be used against the global financial system. Claude Mythos has been shown to identify vulnerabilities in widely used operating systems, raising concern in government and commerce that advanced models may now be able to uncover and potentially exploit cyber-security flaws at a far higher level.
Canada’s Finance Minister, François-Philippe Champagne, said Mythos had dominated discussions at this week’s International Monetary Fund meetings in Washington DC. Anthropic’s developers described the model as “strikingly capable at computer security tasks” after testing for misaligned behaviour. Citing concerns that the model could surface dormant software bugs or identify new ways to exploit weaknesses, Anthropic has chosen not to release it publicly. Access has instead been limited to a small group including Amazon Web Services, CrowdStrike, Microsoft and Nvidia through Project Glasswing, an effort focused on securing critical software. On Thursday, Anthropic released an upgraded version of its existing Claude Opus model so Mythos’s cyber capabilities could be evaluated in less powerful systems.
Independent scrutiny remains limited. The UK’s Artificial Intelligence Security Institute, the only body to have published an independent assessment, said Mythos Preview could compromise systems with weak defences but was not dramatically more capable than Opus 4. Some cyber-security specialists have warned that the alarm may be premature, noting the lack of broad external testing and pointing to OpenAI’s delayed GPT-2 release in February 2019 as a precedent that also drew scepticism.
Senior bankers are now set to receive early access so they can test their own systems before any wider release. C.S. Venkatakrishnan of Barclays said the risks are serious enough to demand closer understanding and rapid fixes. Andrew Bailey said authorities must examine what this latest Artificial Intelligence development could mean for cyber crime, especially if it makes it easier to detect vulnerabilities in core IT systems that bad actors could exploit. The US Treasury has raised the issue with major American banks and urged them to run internal tests ahead of any public release.
For Britain’s small and medium-sized businesses, the risks extend beyond banks themselves because payment, cloud and financial infrastructure underpin daily operations. Anthropic has already said Mythos has uncovered multiple vulnerabilities in core operating systems, financial platforms and web browsers. At the same time, some investors see a commercial opening for the UK technology sector. James Wise of Balderton Capital said Mythos is only the first of many more powerful models likely to expose systemic weaknesses, arguing that companies focused on Artificial Intelligence security and safety could become increasingly important. He said the Sovereign AI unit is a £500m government-backed venture capital fund targeting home-grown Artificial Intelligence businesses.
