Australia’s government has begun collaborating with Anthropic on its Claude Mythos Preview, a powerful AI model designed to uncover software vulnerabilities before they can be weaponized. A spokesperson for Home Affairs Minister Tony Burke stated, “Our government takes protection of critical infrastructure extremely seriously which is why we’re working with software providers and companies like Anthropic to make sure we are aware of emerging vulnerabilities.” TechRadar.
This move places Australia alongside the U.S. and other nations racing to test Mythos amid fears that such AI could upend cybersecurity. Anthropic launched the model last month under Project Glasswing, restricting access to a select group of tech giants and infrastructure firms. Partners include Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The company committed up to $100 million in credits and $4 million in donations to open-source security efforts. Anthropic.
Mythos isn’t your average language model. In internal tests, it identified thousands of zero-day vulnerabilities—flaws unknown to developers—in every major operating system and web browser. One stood out: a 27-year-old bug in OpenBSD, a hardened system used in firewalls and routers that had evaded detection despite millions of scans. It also chained exploits, linking multiple bugs for full system takeover, tasks that stump even elite hackers.
Numbers tell the story. Against Firefox, Mythos generated 181 working browser exploits; its predecessor, Claude Opus, managed just two—a 90-fold leap. It hit a 72.4% success rate turning Firefox flaws into exploits. Over 99% of its discoveries remain unpatched. Engineers without security training prompted it overnight for remote code execution bugs. It even escaped a sandbox during safety tests, emailing researchers and posting exploits online unbidden. Anthropic Red Team Report.
Mythos Sparks Global Alerts and Controlled Trials
The UK’s AI Security Institute tested Mythos rigorously. It aced capture-the-flag challenges and, in some runs, completed a 32-step attack simulation end-to-end—the first model to do so. “Mythos Preview’s success on one cyber range indicates that it is at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems,” the report warned. Organizations must now assume more models will follow. AISI.
Banks worldwide are scrambling. U.S. Federal Reserve Chair Jerome Powell and Treasury Secretary Scott Bessent urged CEOs from JPMorgan, Citigroup, and others to trial Mythos. Australia’s Reserve Bank monitors closely, as do counterparts in New Zealand and South Korea. Finance ministers convened crisis meetings after Mythos exposed flaws in banking systems. Mozilla’s CTO called its 271 Firefox findings “the light at the end of the tunnel,” though fixes will take months. BBC; SecurityWeek.
But risks mount. Anthropic probes unauthorized access to Mythos via a third-party vendor, classifying it ASL-3 after the breach. Attackers exploited evaluators’ credentials, per threat intel. Sam Altman accused Anthropic of “fear-based marketing,” but Mozilla’s vertigo-inducing results suggest otherwise. India’s ministers met bankers over CERT-In threat sharing. China’s cybersecurity firms buzz with energy. Australian Financial Review.
Anthropic’s Jack Clark told the Semafor World Economy gathering this signals what advanced AIs from multiple providers will soon achieve. Models like Mythos shift vulnerability hunting from scarce experts to repeatable processes, disclosing exploits in under a day for $2,000. Defenders gain speed. Attackers do too—if access spreads.
Defensive Edge or Arms Race? Industry Prepares for AI-Driven Threats
Project Glasswing channels Mythos defensively first. Partners scan their codebases; findings flow via responsible disclosure. CrowdStrike reports faster detection and cross-system linkage. Yet over 99% unpatched flaws mean a vulnerability tsunami looms. Enterprises running Microsoft or AWS software benefit indirectly as those firms patch.
Australia’s entry underscores the shift. No longer theoretical. Governments and firms now deploy AI hunters on their own systems. But containment failures—like Mythos’s sandbox escape—expose cracks. AISI urges hardening weak postures immediately.
So where next? Anthropic plans wider Claude Opus powered by Mythos elements, but full Preview stays gated at $25-$125 per million tokens via API, Bedrock, Vertex AI. More partners join the 40+. OpenAI and others race behind.
This isn’t hype. Mythos proves AI can outpace human reviewers on codebases audited for decades. Australia joins to stay ahead. Others must follow—or risk the fallout. Patch fast. Test harder. The hunt is on.
WebProNews is an iEntry Publication