Artificial intelligence (AI) giant Anthropic has expanded access to a highly advanced model deemed too dangerous for public release, including Australia in the select handful of users.

The large language model, known as Claude Mythos, is now being rolled out to an additional 150 organisations across 15 countries, including the Australian government and several local businesses, as part of Project Glasswing.

In an era where large-scale AI launches are happening on a day-by-day basis, this limited, gradual release may seem particularly surprising. But Mythos is not like most other AI systems. Instead it’s an automated tool for assessing software to find critical bugs and vulnerabilities.

This managed release is deliberate, as the discovery of vulnerabilities in computer systems is useful for those who want to defend them and those who want to hack them.

However, the real nature of the impact of AI systems on cybersecurity is significantly more complex.

Finding hundreds of severe vulnerabilities

Under initial testing, Mythos has been able to identify multiple new high-risk vulnerabilities. Left unfixed, such flaws allow attackers to easily steal data or induce system crashes.

While these reports are promising, the raw data needs context. Of the 23,000 vulnerabilities flagged by Mythos, only 6,200 were estimated........

© The Conversation