The Taiwan Banker

What did the financial institutions included in Project Glasswing learn? Was it specific vulnerabilities lurking in their codebases, or more abstract lessons on cybersecurity management? Project Glasswing was a closed-access preview of Anthropic’s Mythos model, claimed to represent a step change in red teaming capabilities. It made headlines by discovering flaws in “every major operating system and every major web browser,” most impressively including a 27-year-old bug in the OpenBSD open-source operating system, which was designed to be security-hardened. The US government has taken a national security view to this capability. On April 7, the same day the Mythos launch was announced, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell met with leaders of major banks. The IMF has also warned about systemic risks to the financial sector. On June 12, the US Commerce Department issued an unprecedented export control order prohibiting access to Mythos (as well as the related Fable model) for any foreign national, prompting Anthropic to temporarily disable the models entirely. As of writing, Project Glasswing remains officially operational, but frozen in practice. These heavy-handed tactics notwithstanding, Mythos-level capabilities will certainly become openly available within a matter of months anyway, or perhaps a year. The project represented a genuine opportunity to clean up infrastructure-level software, which will be a permanent benefit, once patches are in place. A smarter model cannot come along later and discover new logical errors that do not exist – and newer software will be able to verify itself against this new standard. These benefits also have their limits, however, because Mythos comes at a time when the field of cybersecurity was changing to begin with. Vulnerability discovery is no longer the bottleneck at all. Threat actors must be able to exploit the vulnerability, which involves not just the single application, but coordination across layers and systems. In the classic workflow, once a vulnerability was discovered, there would be time to devise and install a patch – but that window has now entirely disappeared. Mandiant, a subsidiary of Google Cloud, has tracked mean time-to-exploit through annual surveys, finding a decline from 63 days in 2018 to negative seven days in 2025: the patch only comes well after the threat has arrived. Patch creation cannot be sped up in the same way that vulnerability discovery can. The former process can mean incorporating business requirements which were not previously formalized, which is more multifaceted than mere automation logic, involving software design in addition to just security. Once the vulnerability is found and the patch is ready, moreover, implementation is not trivial. Financial institutions, like other national security assets, tend to sit higher in the software stack than the infrastructure-level services which earned early publicity for Mythos. For them, the challenge is not directly in fixing their software, but in coordinating with their suppliers (perhaps more than one layer deep), testing patches, and deploying them in stages. To make matters worse, AI also makes it easy to reverse engineer the patch once it becomes public, thereby revealing the original vulnerability. Mythos reportedly can complete this process within 30 minutes. Due to these emerging challenges, the security scholars Heather Adkins, Gadi Evron and Bruce Schneier proposed a need for “VulnOps,” more systematic management of this patching pipeline, in an October blog post. Speaking more broadly about cybersecurity agents, prior to the release of Mythos, they wrote that “vulnerability research could potentially be carried out during operations instead of months in advance,” further reflecting its diminished utility in cyber defense. Thus, for the financial sector, it does not appear that Project Glasswing is mainly about the accumulation of zero-days, if they are using their headstart period wisely. Instead, participants are likely gaining the opportunity to further internalize the “assume breach” mindset. While there is undoubtedly value in the experience of being breached through live red team demonstration, those who were not given the opportunity to participate in the project can still learn the most important lessons. TrendAI, the enterprise business unit of the cybersecurity company Trend Micro, outlined three major principles of the emerging mindset in an April presentation. The first is understanding the context of a vulnerability: most CVE (Common Vulnerabilities and Disclosures) listings are not readily exploitable, even when assigned high ratings. Even those which are, will only be relevant for systems with certain configurations, and organizations with particular risk profiles. Second, not all theoretical exploits are part of active campaigns, so threat intelligence is also necessary to deal with the growing alert volume. Finally, other solutions exist to temporarily mitigate the lack of a patch, such as endpoint detection and response (EDR). An April report from the Cloud Security Alliance entitled The “AI Vulnerability Storm”: Building a “Mythos-ready” Security Program further pointed to insufficient AI automation capability as a critical threat. This triage and response process requires speed and attention beyond the capabilities of humans working alone. At the same time, however, the report also lists the attack surface from unmanaged AI agents as a separate critical threat, highlighting the continuing need for competent oversight. With all these structural changes in the cybersecurity landscape, it is also fair to ask what specifically should be attributed to the one model. Undoubtedly, Mythos has been marketed well. Rather than a sense of foreboding, many cybersecurity practitioners are instead expressing relief that longstanding issues within their field are finally receiving widespread attention. Notably, the OpenBSD bug was found using $20,000 in compute credits – a rate comparable to that found on bug bounty markets. The effects of frontier AI models on cybersecurity cannot be understood without considering economics, in addition to just software engineering. The leading financial institutions were likely already keeping up with the latest developments in cybersecurity, and may also be best able to integrate top models into their workflow. Those who are already behind, on the other hand, will gain only marginal benefit from access to a single model. Much of the fintech sector belongs to this second category, featuring complex system interconnections alongside rapidly iterating codebases. Beyond the economics, however, this episode appears to highlight a persistent weakness in cybersecurity as a profession. Whether or not Mythos truly represents an innovation, the narrative aspect is not just incidental. The ability to tell a stark, yet coherent story to boards and executives is a fundamental skill –which can sometimes be forgotten in the relentless drive to upgrade technical knowledge. Just as “Google” became a verb synonymous with search, Mythos may eventually be associated with a new way of managing cybersecurity – and whether or not that association is fair is beyond the point. The old, static vision of cybersecurity is well and truly dead. Safety by design may still be relevant in certain extreme cases, including vibe coding, or else core low-level logic, such as in the Linux kernel. For real organizations which sit in the middle of the stack, however, resilience however is a fundamentally different matter – a dynamic process which can never be fully completed.