Australian government agencies and businesses are seeking access to Anthropic's Claude Mythos AI model, which the company deems too dangerous for public release [1, 2].
The push for access highlights a growing tension between AI safety and national security. While Anthropic restricts the tool to prevent misuse, Australian officials believe the model is essential for defending against emerging AI-driven threats [1, 3].
Discussions regarding the model, also known as Mythos Preview, have taken place in Canberra and Sydney [1, 2]. The model is considered highly capable for defensive cybersecurity research, which could protect critical infrastructure, and private enterprises from sophisticated attacks [1, 3].
Anthropic has maintained a strict stance on the tool's availability. "We have decided to hold back the full release of Claude Mythos because it is too dangerous for the public at this stage," an Anthropic spokesperson said [3].
The capabilities of the model have already demonstrated significant impact. In one instance, the tool discovered a 27-year-old [4] bug in router software that could affect devices worldwide [4]. This level of vulnerability detection is exactly why Australian agencies are courting the company to establish a local presence or partnership [1, 2].
Despite its intended use for defense, the model's power remains a point of contention. Some reports indicate the model was built specifically for defensive research, while other accounts suggest it may have already been accessed by unauthorized hackers [1, 2].
Efforts to lure Anthropic to Australia involve high-level talks, including those involving Jeffery Bleich, a U.S. envoy, to coordinate AI cooperation between the two nations [1].
“"We have decided to hold back the full release of Claude Mythos because it is too dangerous for the public at this stage."”
The pursuit of Claude Mythos by the Australian government represents a shift toward 'defensive AI' procurement. By attempting to secure a model that is too volatile for the general market, Australia is signaling that the risk of being undefended against AI-powered cyberattacks is now greater than the risk of utilizing a potentially dangerous tool within a controlled government environment.
