AI tool ‘too dangerous to release’ could wreak havoc on businesses

Australia’s largest cybersecurity firm has issued an urgent warning about a powerful new artificial intelligence tool that can find and exploit flaws in software at unprecedented speed and scale, and which experts fear could trigger the next wave of major data breaches.

CyberCX on Thursday told Australian businesses, banks and infrastructure operators they had a closing window to shore up their defences before the technology, or copies of it, reach the hands of criminals.

The warning deals with Claude Mythos Preview, an unreleased AI model built by US company Anthropic that the firm has deemed too dangerous to release publicly. Anthropic has restricted access to about 50 major technology and infrastructure partners – including Microsoft, Google, Apple, Amazon and JPMorgan Chase – under a program called Project Glasswing.

Mythos has already uncovered thousands of previously unknown software flaws, including a 27-year-old bug in an operating system used in firewalls and internet routers worldwide. In one test, it produced 181 working attacks against the Firefox web browser. An earlier, publicly available version managed just two.

Anthropic product lead Angela Jiang said the cyber capability had emerged partly as a by-product of the company’s broader push on coding and long-horizon agentic tasks.

“If something’s really good at coding, it’s also very good at detecting cyberattacks, especially over a bunch of different surfaces, and chaining that together,” Jiang said. The company had been “privileged to work with a bunch of companies to help improve critical infrastructure”.

Dimitri Vedeneev, secure AI lead at CyberCX, said it was not just Mythos’s ability to find long-buried vulnerabilities that set it apart, but also its capacity to chain multiple flaws together and suggest how to exploit them, all from a single prompt.

“Australian organisations should not be waiting for access to Mythos as some kind of silver bullet,” Vedeneev said. “It won’t be long until this capability – or others like it – is more widely available, and potentially in the hands of cyber criminals.”

Mythos has rattled governments worldwide. Bank of England governor Andrew Bailey told the BBC central banks were examining what the technology could mean for cybercrime, while Canadian finance minister François-Philippe Champagne described Mythos as an “unknown unknown” at International Monetary Fund meetings in Washington. The Trump administration convened major US bank bosses to discuss the risks.

Anthropic expects competing AI companies to release similarly powerful tools within 18 months. OpenAI has given a select group of users access to its own cyber-focused model, GPT-5.4-Cyber. Bloomberg reported last week that a small group of unauthorised users had obtained access to Mythos through third parties, and Anthropic has confirmed it is investigating.

Anthropic on Friday separately announced the public beta launch of Claude Security, a defensive product that allows enterprise customers to scan their own code for vulnerabilities and generate patches. It said hundreds of organisations had used the tool in research previews to find flaws “that existing tools had missed for years.” Accenture, Deloitte, PwC, BCG and Infosys are among those using the product.

An Anthropic spokesperson said the company was starting with “some of the largest US-based companies” on the basis that “if they can secure their products quickly, that security extends on a global scale”, adding it looked forward to expanding cybersecurity partnerships.

Not everyone is convinced Mythos represents a clean break. Juraj Janosik, director of AI at cybersecurity firm ESET, said models capable of identifying vulnerabilities had existed well before Mythos. With the right orchestration, threat actors could “already achieve Mythos-like capabilities using generally available models”.

“Many companies still lag in maintaining basic cyber hygiene and are often exploited using older, already public vulnerabilities,” Janosik said. “While the developments in AI capability are a concern, they are dwarfed by the overall lack of cyber resilience.”

Manuel Salazar, director of cyber services at Australian firm Orro, said the fundamentals had not changed. “Mythos changes the speed at which weak fundamentals get exposed,” he said. “For mature organisations, AI is a force multiplier; for less mature organisations, it can accelerate uplift, but it cannot bypass the fundamentals.”

Salazar said Australian businesses should have access to the defensive benefits of Mythos-class AI but not “unrestricted access to a frontier exploit engine”.

“Australia needs to secure access to advanced AI technologies through agencies like the Australian Signals Directorate, Home Affairs and the National Cyber Security Coordinator,” he said. “If we don’t get involved now, we’re at risk of falling behind our US counterparts.”

The Wall Street Journal reported on Thursday that the White House had rejected an Anthropic proposal to roughly double the number of organisations with access to Mythos, citing security concerns. The administration’s relationship with Anthropic has been strained by an earlier dispute over military use of the company’s AI, now the subject of two court cases.

It is unclear whether any Australian organisations are part of Project Glasswing. Anthropic signed an agreement with the Albanese government earlier this year and opened a Sydney office, but no Australian agency has publicly confirmed access.

The warning lands against the backdrop of the 2022 Optus and Medibank breaches, which combined exposed the personal information of millions of customers and reshaped public trust in major institutions. The breaches exploited comparatively conventional weaknesses, and the fear is that more sophisticated AI tools such as Mythos could allow attackers to find and exploit flaws in systems previously considered secure.

CyberCX is urging Australian organisations to map their critical systems, segment their networks and “fight AI with AI” by deploying defensive AI in security functions.

Get news and reviews on technology, gadgets and gaming in our Technology newsletter. Sign up to receive it every Friday.