The April 7 introduction of Claude Mythos was no ordinary software release. Anthropic held back its Mythos Preview, saying that “it reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.”
At the same time, Anthropic organized Project Glasswing, bringing Mythos Preview to “launch partners” ranging from Amazon Web Services to JPMorgan Chase & Co. to Nvidia, and extending access to “over 40 additional organizations that build or maintain critical software infrastructure so they can use the model to scan and secure both first-party and open-source systems.”
Described by Anthropic as “a starting point” for defending cyber infrastructure over coming years, Mythos and Glasswing also marked a turning point in terms of accelerating frontier AI advances; in how private industry will manage them and keep vulnerabilities at bay; and in government involvement and regulatory oversight.
In early June, the previously laissez-faire U.S. administration clamped down on the Mythos 5 model and its constrained Fable 5 version, citing national security concerns.
In Anthropic’s words, federal authorities “issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.” (Access to other Anthropic models was not affected.)
The restriction began to be eased on June 26, with Commerce Secretary Howard Lutnick determining “that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 model,” Semafor reported. Lutnick formally lifted the Mythos 5 and Fable 5 controls, subject to negotiated conditions, on June 30, and Anthropic began restoring access the next day.
Meta, meanwhile, was being pressed to agree – as Anthropic, OpenAI, Google, xAI and Microsoft had already done – to submit models for review by the Center for AI Standards and Innovation (formerly the U. S. AI Safety Institute), according to the New York Times.
OpenAI got clearance on June 26 for a “limited preview” of GPT‑5.6 Sol. CEO Sam Altman said in an X post that working with the government on rolling out models with “significant new levels of capability . . . fits with our long-held strategy of iterative deployment. But this isn’t quite the process that we think is optimal.”
It all may add up to “a shift toward more direct federal involvement in monitoring and controlling advanced AI systems,” risk and compliance advisory firm ACA Group observed. “The development may indicate growing interest among policymakers in establishing more formalized regulatory frameworks and oversight mechanisms, moving beyond ad hoc interventions toward a more structured approach to AI governance.”
Mythos is making waves with financial regulators.
Vice Chair Bowman: “Refining our cybersecurity approach.”
“We know that this model accelerates the process of detecting cyber vulnerabilities,” Federal Reserve Vice Chair for Supervision Michelle Bowman said in a May 1 speech. “On one hand, this capability enables firms to address self-identified vulnerabilities, thereby enhancing cybersecurity. But on the other hand, if used maliciously it could be deployed to identify and exploit weaknesses. As we learn more about this tool and others to be released in the coming weeks and months, we will continue to consider effective supervisory approaches for these and other emerging capabilities.”
Bowman noted that Treasury Secretary Scott Bessent and then-Fed Chairman Jerome Powell had “convened the largest banks to discuss Mythos and the cybersecurity implications of the Mythos model. This type of discussion is extremely beneficial to ensuring the protection of the banking system.”
The Fed vice chair added that “regular communication regarding the unique risks of novel and potentially broadly impactful innovation is necessary. Banks of all sizes have expressed concern about access to the Mythos model. Regulators will continue to focus on critical developments and communicating these risks to supervised institutions, as well as on refining our cybersecurity approach.”
Bank of England Governor and Financial Stability Board Chair Andrew Bailey recognized that Anthropic may have "found a way to crack the whole cyber risk world open,” according to an April 16 Reuters report. At the end of May, Bailey said U.K. banks had not yet worked out access to Mythos. “Spillovers from this sort of cyber risk are so big that we can't just have a single sort of national approach . . . because they're all so heavily interconnected.”
International Monetary Fund analysis “suggests that extreme cyber‑incident losses could trigger funding strains, raise solvency concerns, and disrupt broader markets,” says an IMF blog by Tobias Adrian, Tamas Gaidosch and Rangachary Ravikumar. Claude Mythos’ ability to find and exploit operating-system and web-browser vulnerabilities “foreshadows how fast‑moving, AI‑driven cyber risks could destabilize the financial system if not managed carefully, and why authorities must focus on building resilience through supervision and coordination – rather than treating these developments as purely technical or operational issues.
“On the other hand,” the IMF experts go on, “OpenAI’s specialized, restricted cyber version of GPT‑5.5 assumes vulnerabilities and attacks will grow, and emphasizes equipping defenders more quickly and at scale, under appropriate governance and trusted access models.”
“Frontier AI models can discover zero-day vulnerabilities, which are previously unknown flaws that attackers can exploit before they’re fixed, at a speed and scale that traditional security infrastructure wasn;t built to handle,” says a Deloitte Center for Financial Services article, When Frontier AI Models Outpace Cyber Remediation.
Allison Sagraves: Who takes ownership?
The authors found many institutions were taking “a hard look at their response capabilities by setting up command centers and escalation and triage trees.” They recommended that banks “consider bolstering their cybersecurity response capabilities in four ways: prioritizing which vulnerabilities to address first; improving execution speed for quick remediation; building architectural resilience to reduce risk while simultaneously fixing issues; and implementing governance frameworks that enable faster, more distributed decision-making while maintaining appropriate oversight and control.”
Claude initially “found thousands of serious flaws very fast,” explains Allison Sagraves, who was founding chief data officer of M&T Bank before forming data and AI governance consulting firm Allison Sagraves LLC. “Beyond exposing vulnerabilities, it highlights that most institutions have nothing built to absorb this volume.
“There is no clear owner for how quickly something gets fixed, no way to triage at that volume. Once again we’re at a point where a capability shows up before the governance and ability to operationalize it does.”
The release of a tool like Claude Mythos calls for immediate risk assessment, according to Thomas P. Vartanian, executive director of the Financial Technology & Cybersecurity Center and author of books including “200 Years of American Financial Panics” and “The Unhackable Internet.”
Thomas Vartanian: Careful study required.
“Risk professionals will have to separate reality from hype, which is always present when a new tech product is being marketed,” Vartanian says. “Are Mythos’ new capabilities fact or exaggerated?” Indications are that “Mythos is much more than a brilliant marketing strategy.”
Even if Mythos falls short of the hype or expectations, banks will have to determine whether it or a similar product is required simply to establish under the law a sufficient level of care to be prepared for a significant cyber event. Any tool that can find weaknesses in cyber defenses by definition requires careful study and analysis.
Signaling potential price and performance competition, coming out of China, Z.ai’s GLM-5.2 is said to rival Anthropic’s Fable 5.
To Mark McCreary, chief AI & information security officer, Fox Rothschild LLP, “Mythos is the first model that can autonomously discover zero-days, chain them into multi-step exploits, and do it against compiled binaries without source code. It has been confirmed as the first model to complete a full 32-step simulated corporate network attack. That is a phase shift, not incremental progress.”
The AI may thus be on a revolutionary trajectory, Vartanian suggests, though McCreary has a different angle on it..
“We are accelerating into a threat era that was already forming,” says the Fox Rothschild partner. “Previous models were already finding significant vulnerabilities. Mythos just does it faster, deeper, and with less handholding.
“The fundamentals of good cybersecurity have not changed. The cost of ignoring them has. As one IBM researcher put it, ‘If the attackers aren’t humans anymore, the defenders can’t be humans anymore either.’”
To help identify security vulnerabilities before AI can exploit them, Synack has launched Glasswing-Readiness Assessment. It combines Attack Surface Discovery with Sara (Synack Autonomous Red Agent) “to explore environments at scale and identify real attack paths.”
Synack CEO and co-founder Jay Kaplan said Project Glasswing is “the kind of defensive innovation this moment calls for.” Responding to the challenge “is continuous, agentic AI-driven testing with humans in the loop. Annual assessments tied to a compliance calendar no longer reflect how attacks actually happen.”
“When offensive AI can map an environment and iterate on exploits at machine speed, untested infrastructure like legacy systems, forgotten endpoints, and aging firewalls become the attack surface adversaries find first,” said Dr. Mark Kuhr, Synack co-founder and chief technology officer. “Every weak point is now a viable entry. What looks low-risk in isolation often isn’t once you account for how these attacks actually chain. Getting coverage across your full attack surface is no longer a stretch goal. It's the baseline.”
While vulnerabilities surfaced at Synack’s financial services customers declined 6% in 2025, to 3,215, their severity (critical plus high) climbed from 28.6% in 2024 to 31.4% of the total volume. (Source: 2026 State of Vulnerabilities Report)
Says Vartanian, “Much like Einstein, when he theorized how gravity bends space and time is relative, Mythos and the impact of AI is forcing us to focus on increasingly compressed timelines and a broader field of vision to appreciate the range and scope of cyberattacks and defenses required.
“Changes like these underscore the need for ‘cyber symmetry’: For every AI threat that is developed, there must be AI defenses to neutralize it.”
As the Deloitte Center for Financial Services article pointed out, “These challenges are compounded by the nature of banks’ technology infrastructure, which generally exists as a patchwork: open-source components, third-party platforms, cloud services, and highly regulated transaction systems. This complexity not only creates an extensive attack surface but also makes rapid coordinated response extraordinarily difficult.”
Broadridge Financial Solutions, which is a vendor to a sizable swath of the financial services sector and deploys agentic AI in an “AI partnership model,” joined Project Glasswing in June. “Cybersecurity is fundamental to the resilience of financial markets,” CEO Tim Gokey stated. “We are participating in Project Glasswing to apply frontier AI models to our own systems, helping us stay ahead of emerging threats and supporting a safer financial ecosystem.”
McCreary agrees with others that traditional risk and planning cycles lag machine speed. Risk management – and managers – must be dynamic and adaptive.
Fox Rothschild’s McCreary: “Invest in fundamentals.”
AI will automate much of the lower-level vulnerability assessment that junior analysts handle now. That frees risk professionals to focus on what AI cannot do: business context, risk-tolerance judgment, regulatory implications, and communicating with stakeholders.
“The most valuable risk professionals now are those who can take a finding like ‘this AI model found 23,000 potential vulnerabilities in open-source components we rely on,’ and turn it into a board conversation about capital allocation and regulatory exposure,” McCreary maintains.
Allison Sagraves says organizations should go beyond counting vulnerabilities and measure how fast they are fixed. Remediation speed becomes a key metric.
“Move to continuous security validation,” McCreary advises, “because point-in-time assessments do not cut it when exploit timelines have compressed from months to hours. And invest in the fundamentals: patch management, multifactor authentication (MFA), network segmentation, logging, and access controls.”
In anticipation of evolving regulation, he adds, “Organizations investing in AI-augmented defense today are building the evidentiary record they will need tomorrow. Shorten your patch cycles. Microsoft, Palo Alto Networks and Oracle are all pushing significantly higher patch volumes in response to Mythos findings. If your vendors are patching faster, you need to be deploying faster. And if your incident response plans assume human-speed adversaries, they are already outdated.”