Anthropic PBC’s new artificial intelligence (AI) model, Claude Mythos, has drawn significant attention in the cybersecurity field following its release for a preview by a small group of companies and organizations. According to Anthropic, the latest model is capable of advanced reasoning, which enables it to identify software vulnerabilities.
Based on publicly available information, the preview found Mythos detected high-severity bugs and software defects that had gone undiscovered for decades and are difficult to detect with traditional tools, with vulnerabilities identified across major operating systems and Web browsers.
Mythos was not specifically trained for cybersecurity purposes, but developed as a general-purpose language model offering high performance across a range of uses.
However, the preview showed its ability to identify security vulnerabilities and develop exploits — entirely autonomously and at machine speed. This has raised concerns that malicious actors might exploit such a frontier model to destabilize markets and financial systems. It is not a question of if, but when hackers would use models like Mythos at the expense of others.
Barclays PLC chief executive officer C. S. Venkatakrishnan on Friday warned at a meeting of the G30 consultancy group on the sidelines of the IMF’s spring meeting in Washington that Mythos poses a serious threat to the global banking system and is likely to be followed by even more powerful cyberthreats, Reuters reported.
It followed a surprise meeting earlier this month between US Secretary of the Treasury Scott Bessent, US Federal Reserve Chairman Jerome Powell and several major US bank CEOs to discuss risks posed by Anthropic’s latest AI model, Bloomberg News reported.
Mythos’ limited release to about 40 tech companies and selected organizations has drawn attention, as it allows firms to deploy the model, patch vulnerabilities and bolster defenses while limiting misuse.
It is the first time in the past few years that a leading AI company has withheld a public release due to safety concerns, despite skeptics viewing it as hype ahead of Anthropic’s initial public offering or a marketing strategy.
However, the meeting between Bessent, Powell and bank CEOs — rather than technical or compliance teams — suggested US regulators are taking Mythos seriously and extending AI risk governance into financial supervision.
The preview comes at a time when cyberattack concerns are already a major global issue, as software used in banking systems, medical records, logistics networks and power grids contains vulnerabilities that can be exploited, threatening national security and businesses.
While AI-related cybersecurity risks are not new, Mythos has drawn greater attention due to perceived qualitative changes in AI capability and growing caution about the technology.
While most companies and cybersecurity experts have not yet tested Mythos and cannot fully assess its performance, AI development is clearly advancing faster than expected. There could be a Mythos 2 and Mythos 3 in the pipeline, as new models emerge at an accelerating pace.
Taiwanese authorities and local firms, especially financial institutions, have enhanced defenses through AI-driven detection, vulnerability assessments and upgraded monitoring systems, but traditional cybersecurity measures alone are no longer sufficient.
Addressing these threats requires close coordination between the public and private sectors to harden the domestic cybersecurity ecosystem. Taiwan also needs to engage with like-minded partners on regulatory standards and coordinated countermeasures.
Raising public awareness and improving national preparedness are essential to countering AI-linked cyberattacks, while continuously improving the ability to detect anomalies early, identify threats quickly and respond before incidents escalate remains critical.
