Molotov Attack Redefines AI Executive Security.

TL;DR

• Systemic Cyber Risk Mobilizes Global Institutions: A leaked AI model, Mythos, triggered emergency meetings at three central banks, drawing in the IMF and intelligence agencies. IMF chief Georgieva states, "time is not our friend," highlighting the immediate systemic threat.
• CIA Integrates Agentic AI: Deputy Director Ellis revealed the agency's plan for AI "coworkers" within two years, following the successful generation of its first intelligence report by an AI.
• Anthropic Gains Enterprise Traction Against OpenAI: Ramp AI Index indicates Anthropic is nearing OpenAI in business spending, driven by strong enterprise momentum ahead of a potential October IPO.
• SiFive Secures Capital for RISC-V in AI Infrastructure: Nvidia-backed SiFive raised $400M at a $3.65B valuation, positioning open-architecture CPUs as critical for next-gen AI data centers.
• Polosukhin Launches Secure Agent Runtime, IronClaw: Transformer co-author Illia Polosukhin released a Rust-based secure agent framework, addressing critical credential leakage and sandbox escape vulnerabilities inherent in current agent deployments.

Lead Story: The 2026-04-12 Intel

What began as a late March cybersecurity model leak has now catalyzed a coordinated international response spanning finance, intelligence, and multilateral institutions. Anthropic's Mythos model prompted emergency meetings at three central banks within five days, engaged spy agencies on offensive and defensive applications, and elicited public concern from the IMF.

Fed Chair Powell and Treasury Secretary Bessent convened CEOs from leading financial institutions—Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs—on April 10. The directive was clear: Mythos-class models represent a novel tier of cyber threat, necessitating immediate defensive testing by banks.

Canada's response followed, with the Bank of Canada convening its six largest banks and Desjardins Group. The Bank of England also plans to engage financial institutions within two weeks, involving Treasury, the FCA, and the National Cyber Security Centre.

IMF Managing Director Kristalina Georgieva amplified the concern on CBS Face the Nation, cautioning that the world currently lacks the capacity "to protect the international monetary system against massive cyber risks." Her assessment: "time is not our friend."

Wall Street is not passive. Goldman Sachs, Citigroup, and others are actively testing Mythos internally to assess system vulnerabilities. JPMorgan remains the sole bank publicly identified as a Project Glasswing partner.

The intelligence dimension adds another layer. Defense One reports Anthropic has briefed CISA, NIST, and senior officials across the intelligence community, with NSA analysts already discussing the model. Nextgov highlights the paradox: the administration that previously blacklisted Anthropic from federal contracts is now engaging its most sensitive model for security assessments.

The "Glasswing Paradox" is manifest: the technology capable of exposing fundamental architectural flaws is concurrently being deployed to remediate them. Over 99% of the thousands of zero-day vulnerabilities Mythos identified remain unpatched.

A Christian Science Monitor analysis directly addresses the adversarial landscape. Northwestern's Dr. Subrahmanian warns of formidable Chinese cyber capabilities, suggesting they "probably hacked Anthropic long back." Analyst Allie Mellen underscores the critical risk: "the time between anyone being able to identify and exploit those vulnerabilities is incredibly small."

Some experts still view the fear as overhyped. However, the prevailing concern centers on the certainty that next-generation models from any laboratory will possess similar capabilities, irrespective of design-time guardrails.

In Other News

CIA will integrate AI "coworkers" within two years. Deputy Director Michael Ellis informed Defense One and Nextgov that AI will be embedded across all analytic platforms. The agency has already utilized AI to generate its inaugural intelligence report. Ellis projected that within a decade, officers will supervise teams of AI agents operating as "autonomous mission partners."

Anthropic is narrowing the gap with OpenAI in enterprise spending. Ramp's AI Index shows Anthropic is poised to exceed OpenAI in business spending—a continuation of its February trend where it captured 73% of first-time enterprise AI spend. At HumanX in San Francisco, Anthropic eclipsed OpenAI among 6,500 attendees, demonstrating significant market momentum.

Nvidia-backed SiFive secures $400M for RISC-V AI chips. The open-architecture chip designer achieved a $3.65B valuation in an oversubscribed Series G round led by Nvidia, Apollo, Point72, and T. Rowe Price. SiFive's designs integrate with Nvidia's CUDA and NVLink Fusion, strategically positioning RISC-V as a viable data center CPU for emerging agentic AI workloads.

AI regulation faces bipartisan gridlock. Despite a shared desire to regulate AI, Democrats and Republicans remain divergent on their core concerns—Democrats on corporate accountability and labor impacts, Republicans on issues of censorship and governmental overreach.

X / Social Pulse

The rapid succession of Mythos banking meetings dominated professional discourse. The "three central banks in five days" cadence became a shorthand for unprecedented global regulatory speed, with some noting the political irony of the current administration pursuing legal action against Anthropic while simultaneously urging financial institutions to adopt its model for defensive purposes.

The CIA's "AI coworkers" announcement drew immediate, if varied, reactions. While "First they take our jobs, then they become spies" emerged as a prevalent jest, the concrete detail of an AI authoring its initial CIA intelligence report resonated more profoundly, signaling a tangible shift beyond speculative hype.

Transformer co-author Illia Polosukhin's Business Insider profile, detailing his AI-assisted life while maintaining distrust of autonomous operation, garnered significant interest from security researchers. His IronClaw launch was particularly well-received by developers fatigued by recurrent credential leakage issues in existing agent frameworks.

One to Watch

Agent security is rapidly emerging as a distinct product category. This week underscored this trend: Polosukhin released IronClaw (a Rust-based agent runtime featuring TEE isolation), Anthropic and Nvidia introduced their respective zero-trust agent architectures, and Cisco is reportedly in acquisition discussions with Astrix Security ($250-350M) for agent identity management. RSAC 2026 in late March validated this trajectory: 89% of CISOs are accelerating agentic security adoption. The rate of AI agent deployment is outpacing the development of adequate security tooling.

Quick Hits

• CIA milestone: The agency deployed AI to author its first intelligence report; the Deputy Director positioned it as a "draft aid, not a replacement."
• Claude for Word (beta) launched—offering AI-native drafting and editing functionalities within Microsoft Word, complete with tracked changes. Available for Team and Enterprise plans.
• MiniMax M2.7 shipped, featuring a "self-evolving" training methodology, achieving 56.2% on SWE-Pro and matching Claude Opus 4.6 performance at triple the speed.
• Tech layoffs reached 78,557 in Q1 2026, with 48% attributed to AI automation—marking the highest AI-attributed share on record.
• OpenAI's Ninth Circuit appeal brief remains scheduled for April 30; the Anthropic Pentagon case continues in appellate limbo.

The financial system, the intelligence community, and now the IMF are all actively preparing for a systemic threat before its full manifestation. Whether Mythos signals a true inflection point or merely an overhyped precursor, no significant stakeholder is demonstrating a willingness to defer their response. The incentives have aligned toward proactive risk mitigation.

Sources

Mythos / Glasswing / Banks / IMF: CNBC | Bloomberg | Fortune | Business Insider | Storyboard18 | Bank of Canada/Bloomberg | CBS News (Georgieva) | Just The News

Intel Community / Spy Agencies: Defense One | Nextgov | CSMonitor | NPR

CIA AI Coworkers: Defense One | Nextgov | ChinaTechNews

Anthropic Enterprise / HumanX: Business Insider (Ramp) | The News (HumanX)

SiFive Funding: TechCrunch | HPCwire

AI Regulation: Reason

Agent Security / IronClaw / RSAC: Prism News | Business Insider (Polosukhin) | Dark Reading (RSAC)

Lock in. M. mazen@thorterminal.com