Major U.S. investment banks are evaluating Anthropic's new AI model "Mythos," widely regarded as the most powerful security-focused artificial intelligence ever developed. Despite awareness of the model's dual-use nature, leading banks have decided to leverage Mythos to defend against potentially catastrophic cyberattacks. As competition in AI models intensifies, security threats are growing correspondingly, putting governments and corporations worldwide on high alert.
Bloomberg reported on Sunday, citing sources, that Wall Street banks including Goldman Sachs, Citi, Bank of America, and Morgan Stanley are conducting internal tests of Mythos's security capabilities.
U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell urgently convened chairmen and CEOs of major banks at the Treasury Department in Washington, D.C., on Wednesday, strongly recommending they adopt Mythos to strengthen banking system security.
On the same day bank executives were summoned, Anthropic unveiled Mythos, introducing it as the most powerful model surpassing all existing products. However, Anthropic stated it would only provide the Mythos preview to participants in "Project Glasswing," warning that malicious exploitation by hackers could pose critical risks. Project Glasswing is an initiative through which Anthropic provides $100 million worth of access rights to 50 companies including Amazon Web Services, Microsoft, Google, Cisco, and JPMorgan.
Mythos's distinguishing feature lies in its powerful software flaw detection capability. Anthropic emphasized that Mythos identified most vulnerabilities in major operating systems and web browsers, including thousands of highly dangerous bugs. Notably, it discovered a 27-year-old bug in OpenBSD, an operating system known for robust security, and detected a 16-year-old vulnerability that automated bug-scanning tools had failed to find despite over 5 million scans.
The model also excels at identifying and reproducing cybersecurity vulnerabilities, a capability known as "cyber-gym." In cyber-gym evaluations, Mythos scored 83.1%, far outperforming Anthropic's previous top model "Opus 4.6" at 66.6%. Mythos can simulate attackers with 83.1% accuracy, enabling defenders to precisely detect bugs and create security patches.
Using these capabilities, Mythos successfully executed a simulated attack that exploited vulnerabilities in a corporate networking environment containing misconfigured servers, reused passwords, and outdated software, gaining internal administrator access in one attempt. This marks the first case of an AI model autonomously performing a task that would take human experts more than 10 hours—demonstrating fully automated penetration is possible against companies with weak security.
This means if cyber hacking groups misuse Mythos, it becomes the most powerful attack tool available. The model could identify and reproduce critical flaws in financial institutions or government security systems, potentially causing severe personal data breaches and system paralysis. Anthropic has blocked public access to Mythos due to these risks. The U.S. Treasury appears to have requested Wall Street adopt Mythos, prioritizing defensive capabilities despite acknowledging these dangers.
Simultaneously, the entire Trump administration has been placed on alert. CNBC reported that ahead of the Mythos launch, Vice President JD Vance and Secretary Bessent held private conference calls with major AI company CEOs including OpenAI's Sam Altman, Anthropic's Dario Amodei, Google's Sundar Pichai, xAI's Elon Musk (merged with SpaceX), and Microsoft's Satya Nadella. Vice President Vance intensively reviewed AI model security conditions and cyberattack response measures during the meeting. The Wall Street Journal also reported that the White House, led by National Cyber Director Sean Cairncross, recently inspected conditions at key government facilities, demonstrating that AI risks are a top priority for the Trump administration.
Other nations are also mobilizing. The Bank of Canada discussed responses to Mythos-related cyber threats with major banks on Thursday. The Information reported that Cisco President Jeetu Patel warned last month that AI agents going out of control could become an enormous problem, adding that Cisco is discussing the acquisition of Israeli cybersecurity startup Astrix Security.
