Australian organisations will get access to the artificial intelligence model that Anthropic has deemed “too dangerous” for public release, as part of a staggered global rollout aimed at hardening critical systems against hackers.
On Wednesday morning, the San Francisco company said it had extended access to its Claude Mythos Preview model to about 150 new organisations across more than 15 countries, expanding a program known as Project Glasswing. The Financial Times reported access had been granted across the Five Eyes alliance – which includes Australia – along with NATO and the European Union’s cybersecurity agency.
Mythos has proven extraordinarily adept at finding software flaws that, in the wrong hands, could trigger catastrophic breaches. Anthropic has withheld it from general release, warning its cyber capabilities are dual-use and could enable attacks on financial systems and infrastructure.
A spokeswoman confirmed Australian organisations would gain access but declined to name them. Anthropic has previously briefed Treasury, the Reserve Bank and stewards of critical infrastructure.
The move came a day after Anthropic confidentially filed for a $US1 trillion-plus ($1.4 trillion) initial public offering, and as competitors race to build similarly powerful models within roughly a year.
What is Project Glasswing?
Anthropic launched the program on April 7, initially giving about 50 Wall Street and technology firms access to Mythos to scan their own code for flaws before a broader release could put critical services at risk. The idea is that organisations can use the model to find and fix weaknesses in their systems before attackers do.
The early results were striking. Members surfaced more than 10,000 high- or critical-severity software vulnerabilities since the program launched. The latest cohort opens access to sectors not previously represented, including power, water, healthcare and communications.
“Following several weeks of close collaboration with our Project Glasswing partners, the security industry, open-source software maintainers, and the US government, we’re extending the partnership to approximately 150 new organisations,” Anthropic said in a statement on its website. Many of the new partners are vendors whose codebases are relied upon by governments worldwide.
Which organisations get access?
Anthropic has declined to name the new partners, and such organisations are typically identified only when they choose to disclose it themselves. The original cohort surfaced largely that way, and included Apple, Nvidia, Microsoft, CrowdStrike and Palo Alto Networks, along with Google, Amazon and JPMorganChase. Cloud data firm Rubrik said it was among the new batch of companies.
In Australia, the most likely candidates are systemically important institutions: the major banks, critical-infrastructure operators in energy and telecommunications, and national security and government agencies such as the Australian Signals Directorate and Home Affairs. Financial infrastructure is squarely in scope – JPMorgan was named in the initial US group – making the big four banks a logical fit. Government agencies rarely confirm specific security tooling, so the public may never get a definitive Australian roster.
Why is Claude Mythos Preview considered dangerous?
Anthropic’s most powerful model has proven extraordinarily adept at finding software flaws, and at working out how to exploit them. “What each partner has in common is that a successful attack on their codebase could be catastrophic,” the company said, estimating a major attack could affect more than 100 million people.
The company’s president, Daniela Amodei, said the model was being withheld because its cyber capabilities cut both ways. “Cyber is a dual-use technology … it’s also very good at cyber warfare,” she told Forbes Australia. “We decided that we can’t put this into general availability right now because we believe it’s going to cause a lot of problems in the world if we do.”
The worry, she said, was that a rogue actor could use such a model “for cyber warfare against a state, against a financial institution to hack credit card information, or to try to bring down critical infrastructure”.
Anthropic expects competitors to build similarly powerful models within roughly a year, raising the stakes for defenders to get ahead first.
Get news and reviews on technology, gadgets and gaming in our Technology newsletter every Friday. Sign up here.
David Swan is the technology editor for The Age and The Sydney Morning Herald. He was previously technology editor for The Australian newspaper.Connect via X or email.