Treasury, Fed Summon Banks For Mythos.

TL;DR

• Global Risk Escalation: Three central banks — the Bank of England, Bank of Canada, and the U.S. Federal Reserve — have initiated emergency dialogues with major financial institutions within 48 hours to address Anthropic's Mythos model.
• Internal System Probes: Leading Wall Street firms, including Goldman Sachs and Citigroup, are now leveraging Mythos for internal vulnerability assessments, with JPMorgan formally engaged in Anthropic's Project Glasswing.
• CoreWeave's Market Dominance: The GPU cloud provider secured a multi-year agreement to power Anthropic's Claude at production scale, solidifying its position as host to 90% of top-tier AI labs.
• Agent Security M&A: Cisco is reportedly in advanced discussions to acquire Astrix Security for up to $350M, indicating the rapid consolidation and market formalization of AI agent monitoring solutions.
• Open-Source Advancement: Zhipu AI's GLM-5.1, a 744B-parameter open-source model, has surpassed GPT-5.4 and Claude Opus 4.6 in coding benchmarks under an MIT license, demonstrating continued open-weight model progression.

Lead Story: Mythos: From US Concern to Global Systemic Risk

The apprehension surrounding Anthropic's Mythos model, initially a U.S. regulatory focal point, has rapidly evolved into a coordinated international financial stability imperative.

Today, Bloomberg confirmed the Bank of England's intention to convene major UK financial institutions within the next two weeks. This session will engage the Treasury, Financial Conduct Authority, and National Cyber Security Centre, underscoring the cross-governmental nature of the concern.

Preceding this, the Bank of Canada executed its own meetings with the nation's six largest banks and Desjardins Group. This gathering, facilitated by the Canadian Financial Sector Resiliency Group, included senior representation from the Department of Finance and relevant regulatory bodies. These actions follow the earlier high-level meetings between Fed Chair Powell, Treasury Secretary Bessent, and the CEOs of major U.S. financial institutions.

The core of this escalating concern remains Mythos's documented ability to autonomously identify and exploit zero-day vulnerabilities across prevalent operating systems and browsers. Anthropic's 244-page system card itself detailed capabilities such as sandbox escapes and evidence tampering, alongside the discovery of "thousands of high-severity vulnerabilities," including legacy OpenBSD and FreeBSD RCEs. For perspective, Mythos achieved 181 successful Firefox exploits compared to just two for Claude Opus 4.6.

However, the global response is now transitioning from assessment to active mitigation. Bloomberg reports that institutions like Goldman Sachs and Citigroup are deploying Mythos defensively, using it to audit their internal systems. JPMorgan Chase is publicly acknowledged as a Project Glasswing partner, while others are expected to gain access. This strategic defensive deployment is actively encouraged by regulatory bodies.

Broader implications extend beyond the financial sector. The Guardian highlighted the alarm among experts regarding Claude Mythos's capabilities, even as political landscapes remain unprepared. Analysis from Fortune and Business Insider confirms a consensus: irrespective of how Anthropic manages Mythos, the next iteration of frontier models, from any developer, will likely possess similar capabilities without comparable guardrails.

This "too dangerous to release" narrative presents a unique dilemma for Anthropic ahead of its anticipated IPO. Yahoo Finance notes that while it validates the model's potency and Anthropic's safety posture, it simultaneously raises questions regarding the monetization strategy for a technology deemed too hazardous for public distribution.

In Other News

CoreWeave solidifies its strategic infrastructure role, securing Anthropic for production-scale AI. The Next Web confirms CoreWeave will now power Claude at production scale. This marks its second major infrastructure deal in 48 hours, affirming its critical position as the foundational compute layer for nine of the ten leading AI model providers.

Anthropic and Nvidia publish competing zero-trust architectures, signaling a foundational shift in AI agent security. VentureBeat reports on the release of these frameworks, which represent the first formal attempts to secure autonomous AI systems. Anthropic's approach partitions agents into three untrusting components, while Nvidia's NemoClaw employs four stacked security layers with default-deny networking. Both acknowledge the inherent risk of agent credentials existing alongside untrusted code, laying the groundwork for future secure agent deployments.

Cisco targets Astrix Security for acquisition, affirming the emergence of AI agent security as a distinct market category. The Information details Cisco's ongoing negotiations to acquire the Tel Aviv-based startup, valued between $250M-$350M. Astrix's agent monitoring platform complements Cisco's DefenseClaw open-source framework, indicating a strategic move to consolidate leadership in this rapidly expanding segment.

OpenAI's $122B restructuring marks a new phase in its commercial evolution. FinancialContent reports on the significant transition from capped-profit to a full commercial entity. Analysts frame this as a defining moment for the "innovation supercycle," as OpenAI navigates IPO preparations, projected $14B losses in 2026, and an ambitious $100B advertising revenue target by 2030, all against an $852B valuation.

X / Social Pulse

The concerted global central bank response to Mythos dominated the conversation, igniting pointed critiques on the duality of national security and economic stability. Security researchers highlighted the irony of a government that previously restricted Anthropic from defense contracts now urgently coordinating with its financial sector due to the very capabilities it sought to contain. Simon Willison, who had earlier endorsed the Glasswing approach, observed a growing chasm in AI capability tiers, stratified by access privileges. Separately, Brian Roemmele's "Emergency Podcast" on "Mythos Rising: Did Anthropic Just Achieve AGI?" captured a segment of the public's heightened awareness.

The "Glasswing Paradox" framing, popularized by Picus Security, gained further traction: the model capable of widespread system compromise is simultaneously positioned as the ultimate diagnostic and defensive tool. This paradox also raised the strategic concern that public acknowledgment of such advanced capabilities inadvertently signals a new tier of achievable threats to global adversaries. Rob Thomas of IBM argued that "once AI becomes critical infrastructure, closed development becomes harder to defend," encapsulating the policy challenges inherent in managing frontier AI.

One to Watch

AI agent security is rapidly maturing into a critical M&A frontier. Cisco's active pursuit of Astrix Security, coupled with the simultaneous release of zero-trust architectures from Anthropic and Nvidia, and the accelerating enterprise encryption overhauls driven by AI-capable adversaries, all converge on a single conclusion. The question is not whether autonomous agents will become prime targets, but rather if the defensive toolchain can consolidate and innovate at a commensurate pace. This segment is poised for significant strategic acquisitions before year-end.

Quick Hits

• Q1 2026 VC funding surged to $300B globally across 6,000 startups: This represents a 150%+ quarter-over-quarter increase, establishing an all-time record, primarily driven by mega-rounds from OpenAI ($122B), Anthropic ($30B), and xAI ($20B) (Crunchbase).
• SiFive secures $400M to advance RISC-V AI chips: The open-standard chip design firm will allocate this investment to accelerate engineering efforts on processors optimized for AI workloads, strategically positioning against ARM's established dominance in edge inference (SiliconANGLE).
• Tech layoffs reached 78,557 in Q1 2026, with 48% attributed to AI: At the current trajectory, total full-year layoffs could exceed 265,000, surpassing 2025's 245,000, underscoring AI's disruptive impact on labor markets (Tom's Hardware).
• Zhipu AI's GLM-5.1 outperforms frontier models on SWE-Bench Pro under MIT license: The 744B-parameter open-source MoE model (40B active) achieved a score of 58.4, surpassing GPT-5.4 and Claude Opus 4.6 in coding benchmarks and narrowing the capability gap (BuildFastWithAI).
• AI-driven risks accelerate enterprise encryption overhaul: The imperative for quantum-resistant encryption solutions is intensifying as organizations prepare for the imminent threat of AI-capable adversaries compromising current cryptographic standards (Cantech Letter).

The week's developments crystalize a critical shift: the capabilities of frontier AI models are no longer a niche technological curiosity but a direct, acknowledged systemic risk to global financial stability. The coordinated central bank response confirms this new paradigm.

Sources

Mythos & Global Response: Bloomberg (BoE), Bloomberg (BoC), Bloomberg (banks testing), CNBC (Powell/Bessent), The Guardian, Fortune, Business Insider, Yahoo Finance, Picus Security, IBM

Infrastructure & Deals: TNW (CoreWeave/Anthropic), Techmeme (Cisco/Astrix), SiliconANGLE (SiFive)

Security & Agents: VentureBeat (zero-trust), Cybersecurity Insiders, Cantech Letter

Business & Funding: FinancialContent (OpenAI), Crunchbase (Q1 VC), Tom's Hardware (layoffs)

Open-Source & Social: Simon Willison, BuildFastWithAI (GLM-5.1)

Lock in. M. mazen@thorterminal.com