Anthropic Says Their New AI Is Too Dangerous for Public Use

In a development that signals the accelerating convergence of frontier artificial intelligence and operational risk, U.S. bank executives and regulators held an urgent meeting on Friday to address the cybersecurity implications of Anthropic's newly announced Claude Mythos Preview model.

Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell met with CEOs of the largest U.S. banks to review the emerging threats. The gathering was convened specifically in response to Anthropic's announcement, reflecting the seriousness with which regulators are treating the development.

Anthropic has explicitly chosen not to release Claude Mythos Preview to the public. In its system card and accompanying statements, the company warned that the model can rapidly identify vulnerabilities in virtually every major operating system and web browser using relatively simple prompts. These capabilities, Anthropic noted, could pose severe risks to economies, public safety, and national security if they proliferate beyond responsible actors.

Instead, the company has launched Project Glasswing, a tightly controlled defensive initiative. Limited access to Mythos Preview is being provided to a select group of partners, including Amazon Web Services, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, Palo Alto Networks, JPMorgan Chase, and the Linux Foundation, along with substantial compute resources. The explicit goal: to proactively scan for and remediate critical vulnerabilities across foundational software and critical infrastructure before malicious actors gain comparable tools.

Why This Represents a Material Shift for Financial Institutions

For banks and other financial services firms, the announcement is more than a headline about AI progress. It marks a clear compression in the timeline for sophisticated cyber threats. What historically demanded nation-state-level expertise, time, and resources can now be approximated through advanced prompting by far less sophisticated actors.

Key risks highlighted by the model include:

• Autonomous vulnerability discovery and chaining across core systems relied upon by banking platforms, payment networks, and customer-facing applications.
• Democratization of offensive capability: As cybersecurity experts have noted, the model does not require deep technical expertise to generate actionable exploits, lowering the barrier for a wider range of threat actors.
• Supply-chain and third-party exposure: Many core banking technologies depend on common operating systems, browsers, and open-source components now subject to accelerated scrutiny.

Tech commentator Carmi Levy captured the sentiment succinctly: "This is the kind of thing that should keep us up at night." Cybersecurity specialist Claudiu Popa added that the development is "not a reason to practice fear mongering, but to raise the level of awareness and concern amongst organizations," emphasizing the need to minimize vulnerabilities proactively.

Strategic and Investment Implications

The response from U.S. authorities demonstrates that regulators are treating frontier AI models as a potential systemic operational risk, akin to major cyber incidents or cloud concentration risks that have already shaped supervisory expectations.

For financial institutions and investors, several actionable themes emerge:

Accelerated Defensive AI Investment

Banks are expected to intensify deployment of AI-powered tools for continuous vulnerability scanning, automated patching, red-teaming, and exposure management. Institutions already partnered with Project Glasswing (or able to gain indirect access through vendors) may secure a meaningful first-mover advantage.

Cybersecurity Sector Tailwinds

Demand for specialized platforms in AI-driven threat detection, automated remediation, and secure code analysis is likely to rise. Publicly traded leaders in this space, such as CrowdStrike, Palo Alto Networks, and emerging AI-native security firms, stand to benefit from heightened enterprise and government spending.

Vendor and Supply-Chain Due Diligence

Heightened regulatory and board-level scrutiny of technology providers' own AI risk management practices and patching cadences is probable. This could influence procurement decisions and vendor concentration risk assessments.

Insurance and Capital Considerations

Cyber insurance underwriters and regulators may begin incorporating AI-accelerated threat models more explicitly into pricing, stress testing, and capital adequacy frameworks over the coming quarters.

The Road Ahead

Experts estimate that capabilities comparable to Claude Mythos Preview will reach broader availability within months rather than years. The current period therefore represents a critical window for preparation.

Financial leaders should treat this moment as a strategic call to action rather than a cause for alarm: conduct comprehensive vulnerability inventories, shorten remediation timelines, deepen AI governance frameworks, and explore participation in industry-wide defensive initiatives such as Project Glasswing where feasible.

Anthropic's restrained approach, prioritizing defense over broad distribution, sets a responsible precedent. Yet the underlying capability advance is real and irreversible. The institutions that respond with urgency, technical rigor, and cross-sector collaboration will be best positioned to navigate the evolving AI-driven threat landscape.

In the end, Claude Mythos Preview is not an isolated event. It is an early indicator of how frontier AI will reshape operational resilience across the global financial system. Banks that move decisively today will help define the standards of cybersecurity resilience for the decade ahead.