Pentagon Blacklists, NSA Deploys Mythos.

Mazen Al Sabbagh

— 7 min read
Share

The 2026-04-20 Intel

TL;DR

  • Institutional Bypass: The NSA is deploying Anthropic's Mythos Preview for cybersecurity, leveraging the ODNI's operational lane to circumvent the Pentagon's formal blacklist.
  • Global Regulatory Scrutiny: Major financial bodies—the Bank of England, Fed, ECB, ASIC, and MAS—are now formally monitoring Mythos's systemic banking risks, with industry leaders acknowledging heightened cyber threats.
  • GPT-5.5 Launch Projections: OpenAI's "Spud" model was reportedly detected in production API testing, shifting Polymarket odds to a 76% likelihood for an April 23 release, despite a significant global ChatGPT outage today.
  • Google's Chip Diversification: Google is in discussions with Marvell Technology to co-develop two new AI chips, signaling a strategic move to reduce reliance on Broadcom and enhance inference capabilities.
  • Sarvam AI's Strategic Round: India's foundational model startup, Sarvam AI, is raising $320-350M at a $1.5B valuation, drawing strategic investment from Nvidia and Amazon into a critical, underserved market.

Lead Story: The NSA Is Using Mythos. The Pentagon Blacklisted Anthropic. Both Are True.

Axios reported Saturday evening that the National Security Agency has integrated Anthropic's Mythos Preview model into its cybersecurity operations. This deployment proceeds despite the Department of Defense's official designation of Anthropic as a supply-chain risk. Reuters, TechCrunch, Engadget, and GBHackers have since corroborated the report.

The operational architecture here is critical. The NSA, situated within the Office of the Director of National Intelligence, bypasses direct DoD procurement mandates. This allows a strategic acquisition of Mythos without violating Pentagon restrictions. The tool's access remains constrained to roughly 40 organizations under Anthropic's Project Glasswing. The NSA's choice signals a clear internal assessment: Mythos's cybersecurity utility surpasses the political messaging inherent in the Pentagon's supply-chain risk designation.

This operational divergence significantly weakens the blacklist narrative, an outcome only partially foreshadowed by Friday's "productive" White House meeting between Anthropic CEO Dario Amodei and White House Chief of Staff Jeff Zients. The administration’s position now appears fragmented: blacklisting Anthropic through one agency, hosting its CEO at the highest levels, and simultaneously entrusting its flagship model with national security. When asked about Amodei’s meeting, Trump’s reported response — "Who?" — further underscores the prevailing political incoherence.

Concurrently, financial regulators are formalizing their responses. Reuters today detailed that the Bank of England, the Federal Reserve, the ECB, and the Financial Stability Board are now all officially monitoring Mythos-related banking risks. Australia's ASIC joined this cohort, as did Singapore's MAS, which advised banks to "urgently shore up cyber defences." Deutsche Bank CEO Christian Sewing confirmed European banks are in "close contact" with regulators regarding Mythos. JPMorgan CEO Jamie Dimon, whose bank is a Glasswing partner, acknowledged AI "has made cyber threats worse," even as JPMorgan continues its Mythos evaluations. While the Ninth Circuit May 19 oral arguments provide a legal backstop, the political rationale for the blacklist is increasingly untenable.

In Other News

GPT-5.5 "Spud" Detected in API Testing — Release Window Opens Monday. Unconfirmed reports from Geeky Gadgets and API monitoring services indicate GPT-5.5 was observed in production-scale API testing on April 19. Leaked benchmarks suggest superior performance against Opus 4.7 in coding and 3D simulation. This model, codenamed "Spud," reportedly concluded pre-training around March 24 at Stargate Abilene. Polymarket odds have shifted dramatically, now placing the probability of an April 23 release at 76%, with a broader "by April 30" release at 90%. The final nomenclature — GPT-5.5 or GPT-6 — will depend on the performance delta over GPT-5.4. These reports remain officially unverified, but OpenAI's triple executive departure and Sora shutdown on Friday increasingly suggest a company preparing for a significant launch. The timing is notably juxtaposed with ChatGPT's major global outage today, commencing at 10:05am ET and affecting thousands across the UK and US. The disruption impacted ChatGPT, Codex, and the API platform, with OpenAI still "investigating" hours later. Such an incident presents a difficult optics challenge just days before a potential flagship product launch. Separately, MarkTechPost reported OpenAI is expanding GPT-5.4-Cyber access to additional security organizations today.

Google and Marvell in Talks on Two New AI Chips. The Information reported Saturday that Google is negotiating with Marvell Technology for the co-development of two chips: a memory processing unit and a next-generation inference-focused TPU. This initiative diversifies Google's custom silicon pipeline beyond its established Broadcom partnership. Marvell's stock consequently rose 6.3%. Per Benzinga, the move strategically positions Google to reduce its reliance on Nvidia for inference workloads, a critical cost center as Gemini scales to over 750M monthly active users. Separately, Morgan Stanley published a note arguing that the era of agentic AI will significantly expand CPU demand alongside GPUs, as multi-step autonomous tasks necessitate general-purpose compute beyond what accelerators alone can efficiently provide.

Sarvam AI Raising $350M at $1.5B — India's Frontier Lab Moment. Bengaluru-based Sarvam AI is in advanced discussions to raise $320-350M, with Glade Brook Capital, Nvidia, Amazon, and HCLTech cited as prospective investors, according to the Economic Times. This round would value the company at approximately $1.5B, positioning it as the most valuable non-US, non-China foundational model lab. Sarvam specializes in developing multilingual models optimized for India's linguistic diversity, a problem domain where Western frontier models have historically underperformed. This financing aligns with the broader Q1 2026 trend of record global VC funding, with AI megadeals serving as the primary accelerant.

X / Social Pulse

Musk's xAI demo teased Grok office plugins, showcasing its capability to translate research papers into PowerPoint presentations with direct Excel and Word integration. This, coupled with Simon Willison's recent discovery of Claude's system prompt referencing Chrome, Excel, and PowerPoint agents, solidifies the office-suite AI race into a three-way contest among Anthropic, xAI, and Microsoft Copilot. Today, xAI shipped Grok 4.20 Beta and a Multi-agent Beta to its Enterprise API, alongside the general availability launch of its Speech-to-Text and Text-to-Speech APIs. The latter is aggressively priced at $4.20 per million characters, undercutting OpenAI (~$30/1M) and ElevenLabs (~$50/1M) by roughly 90%. Grok Build and Grok CLI are still anticipated this week, with the 1T-parameter Grok 4.3 model remaining on schedule for completion around April 22.

The Beijing robot half-marathon continues to resonate. NPR, NBC News, The Guardian, and ABC Australia all published significant Sunday analyses. The NYT headline — "A Humanoid Robot Races to a Record Half-Marathon Finish" — understates the profound implications. A 68% year-over-year improvement in race time is not incremental; it represents physical AI's "ChatGPT moment" in terms of public perception and capability demonstration.

The Altman conflict-of-interest narrative is intensifying in the lead-up to the April 27 Musk trial. Startup Fortune published a detailed examination of Altman's stakes in Oklo, Helion, and World Network, raising pointed questions about OpenAI's core beneficiary. Gizmodo reported that investors are now questioning Altman's suitability to lead the company to IPO. Separately, World Network is experiencing distribution bottlenecks for its Orb rollout due to Nvidia chip shortages.

One to Watch

French prosecutors summoned Elon Musk for a voluntary interview in Paris today regarding an investigation into X's handling of AI-generated deepfakes, CSAM, and negationism, as reported by Times of India and The Hindu Business Line. His appearance is uncertain. This summons follows Apple's recent near-ban of Grok from the App Store over deepfake compliance failures and precedes the April 27 Musk-Altman trial by one week. Musk now faces simultaneous legal pressures across the US (OpenAI trial), France (X content moderation), and multiple EU jurisdictions (Grok regulatory actions). The regulatory surface area confronting xAI is expanding at a rate that appears to outpace its legal capacity.

Quick Hits

  • Cursor $50B Round Still in Talks: CNBC confirmed a $2B raise led by a16z and Thrive, with Nvidia as a strategic investor. The company reports $2B ARR, projecting $6B by year-end, with 70% of Fortune 1000 as customers.
  • Stanford AI Index 2026: Follow-up coverage confirms the China-US AI performance gap narrowed from 31.6% to 2.7%, despite the US outspending China by 23x on AI investment, and AI talent migration to the US dropping 89%.
  • Anthropic CEO Dario Amodei: Issued a stark warning that AI could eliminate 50% of entry-level white-collar jobs within one to five years — a significant admission from the CEO whose model the NSA is now running.
  • Maryland Gov. Wes Moore: Scheduled meetings with AI executives to address "Mythos-era threats," indicating a growing parallel in state-level responses alongside federal and international bodies.
  • Anti-AI Violence Trajectory: Continues to draw long-form coverage, with The Independent tracing the escalation "from petitions to protests to petrol bombs," highlighting an intensifying societal backlash.

Sunday briefings rarely deliver this level of market-shaping intelligence. The NSA’s deployment of a model formally blacklisted by the Pentagon represents a profound institutional contradiction, one that typically resolves in a discernible direction. That direction is becoming clearer by the hour. Adding five central banks formalizing Mythos monitoring, Polymarket odds converging on April 23 for GPT-5.5 "Spud," a global ChatGPT outage just days before a potential flagship launch, and Musk’s legal summons in Paris while simultaneously shipping new models, the week ahead is loaded before it even begins. Monday's watch list: the potential GPT-5.5 launch (Wednesday now holds higher market favor), Grok Build/CLI releases, and any official response from the Pentagon regarding the NSA revelation.

Sources

Lock in. M. mazen@thorterminal.com

Read more