AI News Weekly Summary — March 30 – April 5, 2026 | Gemma 4, Pentagon Appeal

Pentagon appeals the Anthropic injunction. Google drops Gemma 4 under Apache 2.0 — ranked #3 globally. Claude Code source leaks via npm error. TurboQuant shakes memory chip stocks. AI recap for March 30 – April 5, 2026.

---

1. Pentagon Appeals Anthropic Injunction — Case Goes to the Ninth Circuit

On April 2, the Department of Justice appealed Judge Rita Lin’s injunction to the US Court of Appeals for the Ninth Circuit. The Pentagon is challenging the order blocking the government’s classification of Anthropic as a “supply chain risk.”

The Ninth Circuit set April 30 as the deadline for the government to file its brief explaining why the injunction should be overturned. Anthropic remains protected by the preliminary injunction — the classification is suspended pending the appeal.

The case moves from district court to the appellate level. How the Ninth Circuit treats the First Amendment argument could define the limits of Pentagon authority over AI companies for years.

2. Google Gemma 4 — Open Model Ranked #3 Globally, Under Apache 2.0

On April 2, Google released Gemma 4 — a family of open-weight AI models built on the same technology as Gemini 3. The key change: an Apache 2.0 license — a first for the Gemma family. Previous versions had a restrictive custom license.

Four variants: E2B (2.3B effective parameters), E4B (4.5B effective), 26B MoE (4B active), and 31B dense — the latter ranked #3 globally on Arena AI ELO (1,452 pts), above many closed models with far more parameters.

Context window up to 256K tokens, native vision and audio, 140+ languages. Models run on consumer hardware — from smartphones to workstations.

Apache 2.0 is the real revolution — it removes the barriers that kept enterprises from deploying Gemma in commercial products. Google is effectively giving away frontier-level technology for free.

3. Claude Code — Full Source Code Leaked via npm Packaging Error

On March 31, Anthropic accidentally published the full source code of Claude Code (~500K lines of TypeScript, ~1,900 files) in an npm package. The cause: a misconfigured source map file in version 2.1.88 — a 59.8 MB .map file exposed the entire codebase, and a publicly accessible R2 bucket allowed source downloads.

Anthropic confirmed this was “a human error in the release process,” not a security breach. The repository hit over 25,000 stars on GitHub within hours.

4. npm Supply Chain Attack — Trojan in the axios Package

On the same day (March 31), independently of the Claude Code leak, an attack hit the axios npm package. Two malicious versions (1.14.1 and 0.30.4) contained a Remote Access Trojan (RAT) connecting to a server operated by the Sapphire Sleet group.

The attack window ran from 00:21 to 03:29 UTC — before npm pulled the malicious packages. Axios is a widely used HTTP library depended on by thousands of projects, including Claude Code. Anyone who installed or updated packages depending on axios during that window may have downloaded the infected version. The Claude Code leak itself occurred later (~04:00 UTC), after the trojan was already removed. Microsoft published a mitigation advisory.

Two events on one day — a source code leak and a supply-chain attack — serve as a warning for the entire npm ecosystem. Even the best-funded AI companies are vulnerable to basic configuration mistakes.

5. Google TurboQuant — 6x Memory Compression, Chip Stocks Drop

Google Research published TurboQuant — an algorithm that compresses LLM key-value (KV) caches down to 3 bits with no accuracy loss. On NVIDIA H100 GPUs, it achieves an 8x speedup in attention logit computation and at least 6x reduction in KV cache memory usage.

The algorithm requires no additional training. Presentation at ICLR 2026.

Market reaction: Micron and Western Digital stocks dropped after publication — analysts say that if AI companies can cut memory requirements 6x through software alone, HBM demand may be lower than projected.

The market fears the Jevons paradox won’t apply this time — “more efficiency = fewer chips” instead of “more efficiency = more demand.” If those fears hold, Samsung and SK Hynix will feel it first.

6. Anthropic Mythos — Early Access Testing, Cybersecurity Warnings

Anthropic continues testing its Mythos model (internal name: Capybara) with a group of early-access customers. The company has not set a public release date — partly because the model remains expensive to run at scale.

Key detail: Mythos “plans and executes sequences of actions on its own, moving across systems, making decisions, and completing operations without waiting for human input at each stage.” Anthropic is privately warning governments the model poses “unprecedented risks” in cybersecurity.

Fortune: Mythos “dramatically outscores Claude Opus 4.6 in coding, reasoning, and cybersecurity.”

7. Mistral AI — $830 Million in Debt for a Data Center Near Paris

On March 30, Mistral closed $830 million in debt financing — the largest such raise by a European AI startup. A consortium of 7 banks (Bpifrance, BNP Paribas, Crédit Agricole, HSBC, La Banque Postale, MUFG, Natixis).

The funds will build a data center in Bruyères-le-Châtel near Paris: 13,800 NVIDIA GB300 GPUs, 44 MW capacity. Opening: end of June 2026. Mistral’s target: 200 MW total capacity across Europe by end of 2027.

8. Eli Lilly and Insilico Medicine — $2.75 Billion AI Drug Discovery Deal

On March 29, Eli Lilly expanded its partnership with Insilico Medicine (Hong Kong) in a deal worth up to $2.75 billion. Lilly pays $115 million upfront; the remainder depends on regulatory and commercial milestones plus royalties.

Lilly gets exclusive global rights to develop and commercialize drugs created on the Pharma.AI platform. Insilico has developed at least 28 drugs using generative AI, with nearly half already in clinical stages.

This is the largest AI-assisted drug discovery deal in history. A signal that big pharma treats AI not as an experiment but as a core R&D strategy.

9. Oracle — 30,000 Workers Get a 6 AM Layoff Email

On April 1, Oracle executed the mass layoffs it had been signaling for weeks. According to sources, up to 30,000 employees in the US, Canada, Mexico, India, and other countries received termination notices via email at 6 AM local time.

The layoffs fund $56 billion in AI infrastructure CAPEX. Oracle stock rose on the cuts.

10. Tennessee — First State to Ban AI Therapy Chatbots

Tennessee Governor Bill Lee signed SB 1580 into law — banning deployment of AI systems that represent themselves as “qualified mental health professionals.” Effective July 1, 2026.

Votes: Senate 32-0, House 94-0 — full bipartisan support. The law includes a private right of action — patients can sue chatbot developers.

Separately, Tennessee is processing SB 1493 — a bill making it a felony to train AI to encourage suicide.

11. UK — Regulatory Framework for Agentic AI

On March 31, the UK’s DRCF (Digital Regulation Cooperation Forum) published “The Future of Agentic AI” — a five-level autonomy spectrum for AI systems, from simple tools to fully autonomous agents.

Authors: four key regulators — CMA (competition), FCA (finance), ICO (data protection), Ofcom (media). Identified risks: algorithmic collusion, action bundling, prompt injection, data minimization failures.

The paper concludes that existing UK legal frameworks cover agentic systems — no new legislation needed, but cross-regulator coordination is required.

12. SpaceX Files Confidentially with SEC — IPO on the Way

On April 1 — despite the date’s association with pranks — SpaceX filed a confidential S-1 with the SEC, preparing for what could be the largest IPO in history. The filing was confirmed by Bloomberg and Reuters.

Following its February acquisition of xAI (combined valuation: $1.25 trillion — SpaceX $1T, xAI $250B), SpaceX positions itself as an “integrated innovation engine” combining AI, rockets, and satellite internet. Analysts estimate IPO valuation at up to $2 trillion.

13. NVIDIA — National Robotics Week and Physical AI Data Factory Blueprint

During National Robotics Week, NVIDIA announced the Physical AI Data Factory Blueprint — an open-source project (GitHub, April 2026) for building data pipelines for robotics, vision AI agents, and autonomous vehicles.

NVIDIA highlights advances in transferring robots from simulation to the real world — models trained in virtual environments achieve increasingly reliable zero-shot transfer.

14. In Brief

• Rebellions (South Korea) closed a $400 million round to develop inference-optimized AI chips — an NVIDIA GPU alternative targeting the Asian market.
• Runway launched a $10 million fund and Builders program for startups at the intersection of AI, media, and world simulation. Selected companies receive free API credits.
• Google Workspace added Vids (text-and-slide-to-video generation), Lyria (AI music creation from prompts), and Veo (advanced AI video with frame-by-frame editing). All tools available directly in the Workspace editor — no app switching required.

15. Macy’s — AI Shopping Assistant Boosts Spending 4.75×

Macy’s quietly launched an AI shopping assistant powered by Google Gemini. Test data: customers using the assistant spent 4.75× more per visit than others. Internal launch: December 2025, public: March 23, 2026.

16. GPT-5.4 — New Benchmarks Confirm Computer Use Dominance

GPT-5.4 (OpenAI, released March 5) — this week, independent benchmark results confirmed the model’s record-setting performance. On OSWorld-Verified and WebArena Verified, it officially crossed 75% — human-level on desktop computer tasks. On GDPval (knowledge work tests): 83% — a record.

The GPT-5.4 Thinking variant integrates test-time compute, combining reasoning, coding, and agentic workflows in a single model.

17. Grok 4.20 — Top Score on News and Reasoning Benchmarks

xAI released Grok 4.20 — built on the Grok 4 architecture with extended real-time web access. The highest-scoring March release on benchmarks measuring accuracy against news and events from the past 30 days. On MMLU-Pro: 89.2%, on LiveBench News: 94.1%. Available to X Premium+ subscribers and via the xAI API.

18. Tech Layoffs Q1 2026 — Over 52,000 in 3 Months

Challenger, Gray & Christmas data: the tech sector lost 18,720 jobs in March, bringing Q1 2026 total to 52,050. Including April (Oracle’s 30K): the number rapidly approaches 90,000+ for the year.

The trend: companies increasingly pair layoff announcements with AI investment announcements. Meta, Oracle, Block, Atlassian — each cites AI as the reason or co-reason for cuts.

---

Weekly Takeaway

The week of March 30 – April 5, 2026 moved the Anthropic vs. Pentagon saga to the appellate level — the Ninth Circuit will review the case with an April 30 deadline. Google released Gemma 4 under Apache 2.0 — an open model ranked #3 globally, a signal that the open-source AI race is intensifying. Claude Code lost its source code through a basic npm error, while a parallel supply-chain attack on axios exposed the fragility of the ecosystem. Google’s TurboQuant shook memory chip stocks — 6x lower memory requirements through software alone. Mistral took $830M in debt for a Paris data center. And Oracle laid off up to 30,000 people with a 6 AM email.

---

Sources: Axios, CNBC, CNN, ABC News, Bloomberg, Fortune, TechCrunch, Reuters, Washington Post, VentureBeat, Tom’s Hardware, The Hacker News, Malwarebytes, Microsoft Security Blog, SANS Institute, Huntress, Google Blog, Google Research Blog, Google Open Source Blog, NVIDIA Blog, Defense One, InsideDefense, The Hill, Motley Fool, STAT News, Fierce Biotech, Pharmaceutical Technology, SiliconANGLE, Transparency Coalition, Troutman Privacy