The Bank of England, the Financial Conduct Authority, and the UK Treasury are engaging in urgent discussions with the National Cyber Security Centre to assess the risks posed by Mythos Preview, Anthropic’s latest AI model. In the United States, Treasury Secretary Scott Bessent held a similar meeting with major Wall Street banks. This marks the first time that financial regulators are coordinating a response to the offensive capabilities of an AI model.
Mythos, Anthropic’s newest model, raises significant alarms due to its unprecedented offensive cybersecurity capabilities within the realm of generative AI. The model can autonomously combine multiple vulnerabilities to construct complex exploits; in one documented case, it produced a browser exploit incorporating four different flaws, utilizing escape techniques from both the rendering engine and the operating system. During just a few weeks of internal testing, Mythos identified thousands of critical vulnerabilities, including zero-day flaws across every major operating system and web browser, as well as a 27-year-old bug in OpenBSD, a system known for its robust security. The systemic risk is compounded by the inherent nature of these capabilities. “They emerged as a direct result of overall improvements in code, reasoning, and autonomy,” Anthropic stated in the announcement of Project Glasswing, clarifying that the model was not deliberately trained for these offensive purposes. This is precisely why Anthropic has opted not to commercialize Mythos Preview, a decision parallel to OpenAI’s choice in 2019 concerning GPT-2, albeit for different reasons back then.
Anthropic decided against bringing Claude Mythos Preview to market, citing its offensive cybersecurity features as incompatible with general availability. Only OpenAI had made a comparable decision back in 2019 with GPT-2, but for different reasons. In the short span of internal testing, Mythos Preview uncovered thousands of critical vulnerabilities, including zero-day flaws in all major operating systems and web browsers, along with a 27-year-old bug in OpenBSD, renowned for its security. Anthropic elaborated that these offensive capabilities were not a product of targeted training. “They emerged as a direct result of overall improvements in code, reasoning, and autonomy,” the company detailed during the unveiling of Project Glasswing.
Project Glasswing: A Defensive Coalition of 12 Partners
The dual nature of these capabilities was immediately recognized by financial regulators. A model that can chain together four vulnerabilities to create a browser exploit, gain local privileges on Linux through race condition exploitation, or craft a remote code execution exploit for the NFS server on FreeBSD represents a qualitative shift in the threat landscape. The Bank of England, the FCA, and the Treasury are assessing whether critical financial sector computer systems—such as trading infrastructures, payment systems, and insurance platforms—contain vulnerabilities that could be targeted by malicious actors with capabilities similar to those of Mythos Preview.
In response to this risk, Anthropic launched Project Glasswing, a defensive coalition comprising 12 founding partners and over 40 additional organizations tasked with maintaining critical software infrastructures. Founding partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Each partner is granted access to Claude Mythos Preview to detect and rectify vulnerabilities in their own systems, as well as in the open-source code that underpins much of the global infrastructure. Anthropic is committing $100 million in usage credits for the model throughout this research phase, along with $4 million in direct donations to open-source security organizations, including the Linux Foundation, OpenSSF, and the Apache Software Foundation. Following the free phase, access to Mythos Preview will be billed at $25 per million input tokens and $125 per million output tokens, available through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
Patch Timelines Incompatible with AI Speed
Banking, market, and insurance infrastructures are structurally exposed to the risks highlighted by Mythos Preview. Interconnected legacy systems and tight patching timelines dictated by operational continuity obligations create prolonged exposure windows. Unfortunately, these windows have shrunk drastically. “The window between discovering a vulnerability and its exploitation by an adversary has collapsed from several months to mere minutes with AI,” stated CrowdStrike, a founding partner of Project Glasswing.
For Chief Information Security Officers in the financial sector, the challenge extends beyond patching known vulnerabilities. Mythos Preview has demonstrated its ability to identify novel vulnerabilities that lack representation in existing CVE databases. As a result, security teams can no longer rely on standard threat intelligence feeds to anticipate attack vectors that malicious actors—equipped with similar capabilities—might exploit.
A Debate on Anthropic’s Communication Strategy
The regulatory response has sparked debate within the tech community. Numerous experts challenge the unique nature of Mythos Preview’s capabilities, arguing that smaller open-source models replicate a significant portion of the analyses presented by Anthropic. The AI vulnerability research firm AISLE claims to have achieved comparable results on several of the cases through less expensive models, asserting that the competitive edge lies in the tooling rather than the model itself. David Sacks, a technology advisor at the White House, has publicly expressed concerns regarding the credibility risk for Anthropic if the threats they outlined fail to materialize. This tension between the need to alert defenders and the risk of rhetorical overreach presents an important aspect of AI governance that CIOs and CISOs must consider when interpreting announcements from leading frontier model providers.
Coordination between London and Washington occurred in less than a week, demonstrating that financial authorities now treat AI-related risks as an active threat. For IT and security leadership in European financial institutions, the pressing concern revolves around when hostile actors may gain access to capabilities comparable to those of Mythos Preview. Newton Cheng, head of cybersecurity Red Team at Anthropic, conveyed this clearly: “The capabilities of frontier models are expected to advance significantly in the coming months.”



