The AI Dispatch

Cybersecurity & Financial Regulation

Powell and Bessent Summon Bank CEOs to Emergency Briefing on Mythos Cyber Risk

Federal Reserve Chair Jerome Powell and Treasury Secretary Scott Bessent convened an urgent meeting with the heads of America’s largest banks to warn that AI-driven cyberattack capabilities may have outpaced the financial system’s defenses.

Sources: Bloomberg • CNBC • Fortune

The Federal Reserve Chair does not summon bank CEOs on a Thursday morning unless the threat is already inside the building. Jerome Powell and Treasury Secretary Scott Bessent did exactly that this week, calling the chief executives of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo to an emergency cybersecurity briefing focused on Claude Mythos — the Anthropic model whose autonomous capabilities have dominated headlines since its restricted release. JPMorgan’s Jamie Dimon was the only major bank chief absent from the meeting.

The briefing’s core message, according to people familiar with the discussion, was blunt: current financial-sector cyber defenses were designed for human-speed adversaries and may be structurally inadequate against AI systems capable of identifying and exploiting vulnerabilities at machine speed. Regulators walked attendees through classified threat assessments describing potential attack vectors — from automated social engineering of bank employees to real-time exploitation of zero-day vulnerabilities in trading infrastructure — that Mythos-class models could theoretically execute with minimal human oversight.

The meeting signals a new phase in the government’s response to frontier AI capabilities. Where previous regulatory actions focused on model access and export controls, this intervention targets the downstream institutions that would bear the consequences of an AI-enabled breach. Several attendees reportedly left the session requesting follow-up technical briefings from their own security teams, and at least two banks have since initiated emergency reviews of their AI-facing threat models.

Export Controls & Geopolitics

DeepSeek Accused of Training V4 on Smuggled Nvidia Blackwell Chips

Reports allege the Chinese AI lab sourced export-banned Nvidia Blackwell GPUs through a smuggling network to train its next-generation model, reigniting fierce debate over whether U.S. chip controls are enforceable.

Sources: The Information • Tom’s Hardware

The Information reported this week that DeepSeek trained its upcoming V4 model on Nvidia Blackwell-architecture GPUs — chips that have been explicitly banned from export to China under U.S. Commerce Department rules since late 2024. According to the report, the hardware was sourced through an intermediary smuggling ring and is physically located in a data center in Inner Mongolia. Nvidia issued a statement calling the claims “far-fetched,” but declined to comment on specific supply-chain tracking measures.

The allegations strike at the heart of Washington’s semiconductor containment strategy. The entire logic of export controls rests on the assumption that physical chokepoints in the chip supply chain can effectively throttle adversarial AI development. If Blackwell-class hardware can reach Chinese labs through gray-market channels at training-relevant scale, the policy achieves little beyond raising costs and alienating allied chip buyers who face the same licensing bureaucracy. Congressional hawks are already calling for expanded enforcement powers, while industry voices argue the episode proves that supply-side controls are a leaky vessel for managing a demand-side problem.