Australia Gains Access to Anthropic's Mythos AI
ABC News reports that Australia's national security agencies and critical-infrastructure defenders are among about 150 organisations across more than 15 countries newly granted access to Anthropic's Claude Mythos Preview, as part of an expansion of Project Glasswing. Anthropic has withheld the model from broad public release because of its cybersecurity capabilities; it can identify software and network vulnerabilities that would pose serious risk if misused. ABC reports access was previously limited to the UK and US governments and about 50 US companies, and that the expanded group spans allied countries including Canada, Germany, Japan, India, and South Korea, with each organisation required to meet security requirements first. Australia's federal government declined to confirm details of its access while its cyber security agency welcomed the inclusion. ABC notes independent MP Kate Chaney urged faster government action on AI, and cites a report warning that AI-driven data-centre demand threatens Australia's energy transition.
What happened
ABC News reports that Anthropic has granted about 150 organisations across more than 15 countries, including Australia, access to its Claude Mythos Preview model as part of an expansion of Project Glasswing. According to ABC, Anthropic has not released the model publicly because of its cybersecurity capabilities, and it can identify software and network vulnerabilities that would pose significant risk if misused. ABC reports access was previously limited to the UK and US governments and about 50 US companies.
Who gets access
ABC reports the expanded group spans countries allied with the United States, including Canada, Germany, Japan, India, and South Korea, with each organisation required to meet security requirements before gaining access. Australia's federal government declined to confirm details of its access, while the national cyber security agency welcomed the inclusion.
Technical and policy context
A model that can automatically surface exploitable code and network weaknesses is a clear example of dual-use AI, offering defensive benefits for vulnerability discovery alongside potential for offensive misuse. For practitioners, that raises operational questions about access controls, red-team practices, and secure deployment when integrating vulnerability-detection capabilities into defensive workflows.
What to watch
Open threads include whether agencies or companies disclose usage policies and oversight for restricted-model access, any technical detail from Anthropic on deployment constraints, and follow-up on the cited report warning that rising AI data-centre demand could strain Australia's energy transition. Independent MP Kate Chaney has urged the government to accelerate its response to AI risks and opportunities.
Key Points
- 1Anthropic expanded access to its withheld Claude Mythos Preview model to about 150 organisations across 15-plus countries, including Australia's cyber-defense agencies, under Project Glasswing.
- 2The model is restricted because it can find software and network vulnerabilities, a dual-use capability with both defensive value and misuse risk; each organisation must meet security requirements first.
- 3Editorial analysis (generic industry): staged access to a high-capability cyber model pressures governments and procurers to update oversight, controls, and incident-response practices.
Scoring Rationale
Extending a deliberately withheld, cyber-capable frontier model to roughly 150 organisations across 15-plus countries, including national security agencies, is a significant security and governance development with direct relevance to defenders, procurement, and policy. Coverage across Anthropic's own announcement and outlets including TechCrunch, Cybersecurity Dive, and CyberScoop supports the weight, though it is an access expansion rather than a new model release.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
