Callista AI Weekly (March 23-29, 2026)

The week's defining story was a leak. Anthropic accidentally exposed internal documents revealing a next-generation model called Claude Mythos, described internally as "a step change" in reasoning. But the leak was only one thread in a week that also brought a federal judge blocking the Trump administration's Anthropic ban, OpenAI shutting down Sora, Arm shipping its first-ever silicon product, and a new AI benchmark that broke every frontier model.

New AI Use Cases

Harvey Hits $11 Billion

Harvey, the legal AI startup, closed a $200 million round on March 25, pushing its valuation to $11 billion. Sequoia and Singapore's GIC co-led the raise. The four-year-old company now serves more than 100,000 lawyers across 1,300 organizations, with annual recurring revenue reaching $190 million in January - nearly double the $100 million it reported last August. Harvey's tools handle contract analysis, compliance checks, due diligence, and litigation support. Its valuation has jumped more than 3.5x in a single year.

domo.health Unveils Voice-First Care Agent in Zurich

Swiss healthtech company domo.health presented a voice-first AI agent for homecare professionals at the Google Healthcare & Life Sciences Community 2026 event in Zurich on March 26. The agent, built on Google Cloud's Gemini Live API and Vertex AI, is integrated into domo.health's PRO platform. It helps homecare staff document care, access patient records, and coordinate workflows through natural conversation - reducing the administrative overhead that pulls time away from patients.

Cisco Launches Security Tools for the Agentic Workforce

At RSA Conference 2026, Cisco rolled out a suite of security tools designed for enterprises deploying AI agents. The centerpiece is DefenseClaw, an open-source secure agent framework that automates security inventory and integrates with NVIDIA OpenShell. Cisco also introduced new Splunk AI capabilities that automate response workflows for security operations teams. The timing is pointed: in a recent Cisco survey, 85% of enterprise customers said they were experimenting with AI agents, but only 5% had moved them into production. The gap between experimentation and deployment is a security problem, and Cisco is betting it can own that space.

Arm Ships Its First Chip, Meta Signs On

In a historic first, Arm announced on March 24 that it will sell its own silicon product after 35 years as a chip design company. The AGI CPU packs up to 136 Neoverse V3 cores on TSMC's 3nm process, claims more than 2x performance per rack versus x86 processors, and targets the rising demand from agentic AI workloads. Meta is the lead customer and co-development partner, planning to deploy the chip alongside its custom MTIA accelerators across gigawatt-scale data center infrastructure. Other launch partners include Cerebras, Cloudflare, OpenAI, and SAP. Arm's stock jumped 16% on the news.

Major Vendor Updates

OpenAI Kills Sora, Launches Foundation

OpenAI made two very different moves on March 24-25. First, the shutdown: OpenAI confirmed it will discontinue Sora, its AI video generation platform, including the iOS app, API, and Sora.com. The app closes April 26, and the API sunsets September 24. The numbers tell the story. Sora was burning roughly $15 million per day in inference costs at peak, while total lifetime revenue from in-app purchases reached just $2.1 million. Downloads fell from 3.3 million in November 2025 to 1.1 million by February 2026. A three-year deal with Walt Disney Co., announced in December 2025 with plans for $1 billion in AI video investment, is being wound down.

The video generation model itself survives inside ChatGPT for paying subscribers, but the standalone product ecosystem is gone. For developers who built on the Sora 2 API, it is a hard deadline.

One day later, Sam Altman announced the OpenAI Foundation with a pledge to grant $1 billion over the next year. The funding targets life science research - including Alzheimer's and high-burden diseases - along with programs to mitigate AI's impact on jobs, the economy, and children's mental health. "AI will also present new threats to society that we have to address," Altman wrote. "No company can sufficiently mitigate these on its own."

Anthropic's Mythos Leak and Legal Victory

It was a turbulent week for Anthropic. On March 26, security researchers discovered nearly 3,000 unpublished documents in a publicly accessible data store, including draft blog posts for an unreleased model called Claude Mythos (also internally codenamed Capybara). A misconfigured content management system had automatically made uploaded files public.

The leaked drafts describe a model that gets "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity" compared to Claude Opus 4.6. They also warn that the model could significantly heighten cybersecurity risks by rapidly finding and exploiting software vulnerabilities. Anthropic confirmed the model is real, calling it "a step change" in capabilities, and said it is currently being tested with a small group of early access customers. The company acknowledged the model is "very expensive for us to serve, and will be very expensive for our customers to use."

Separately, on March 26, U.S. District Judge Rita F. Lin issued a preliminary injunction blocking the Trump administration's ban on government use of Anthropic's technology. In a 43-page ruling, she struck down the Pentagon's designation of Anthropic as a "supply chain risk to national security," calling it "classic First Amendment retaliation." The dispute originated from two restrictions in Anthropic's usage policy: Claude may not be deployed for fully autonomous weapons systems or for mass domestic surveillance. Judge Lin wrote that "the record supports an inference that Anthropic is being punished for criticizing the government's contracting position in the press."

ARC-AGI-3 Breaks Every Frontier Model

ARC Prize launched ARC-AGI-3 on March 25 at Y Combinator HQ in San Francisco, with a fireside chat between creator Francois Chollet and OpenAI CEO Sam Altman. The new benchmark is the first fully interactive test in the ARC-AGI series, featuring hundreds of turn-based environments handcrafted by human game designers. There are no instructions, no rules, and no stated goals. To succeed, an AI agent must explore, learn, and carry knowledge across increasingly difficult levels.

The results so far are stark. Humans score 100%. The best-performing frontier model, Gemini 3.1 Pro, scored 0.37%. GPT-5, Claude, and Gemini all came in below 1%. The competition carries more than $2 million in prizes, and all winning solutions must be open-sourced.

MCP Crosses 97 Million Downloads

The Model Context Protocol (MCP), originally introduced by Anthropic in November 2024, hit 97 million monthly SDK downloads in March 2026. That is comparable scale to React, which took roughly three years to reach 100 million monthly downloads. MCP did it in 16 months. The ecosystem now includes more than 5,800 community and enterprise servers. Adoption accelerated after OpenAI committed to MCP support, breaking the provider-specific tool format fragmentation that had slowed agentic AI development. Google DeepMind and Microsoft have also adopted the protocol.

Cursor-Kimi Attribution Controversy

Cursor sparked a credibility crisis this week when developers discovered that Composer 2, its latest coding model, was built on Moonshot AI's open-weight Kimi K2.5 model. Cursor had marketed Composer 2 as a proprietary advance, highlighting "continued pretraining" and "scaled reinforcement learning" without mentioning Kimi K2.5 in any of its public materials. A developer found the Kimi model identifier in Cursor's API responses.

Cursor's VP of developer education Lee Robinson acknowledged that "Composer 2 started from an open-source base," estimating that about 25% of compute came from the base model. But the Kimi K2.5 license requires prominent attribution for any commercial product exceeding $20 million in monthly revenue - a threshold Cursor surpasses by roughly 8x. Moonshot AI confirmed Cursor used the model "as part of an authorized commercial partnership," but the episode has raised questions about transparency in the growing practice of fine-tuning open models for commercial products.

AI Governance

White House Releases National AI Policy Framework

On March 20, the White House published a National Policy Framework for Artificial Intelligence, outlining legislative recommendations for Congress. While the document itself creates no new legal obligations, its policy direction is clear: broad federal preemption of state AI laws, reliance on existing sector-specific regulators rather than a new federal AI authority, and regulatory "sandboxes" to accelerate AI deployment.

The framework would prohibit states from regulating AI model development or imposing liability on AI developers for third-party misuse. It carves out exceptions for state authority over child protection, fraud prevention, and consumer safety. Federal preemption has failed twice in this Congress - it was removed from the GOP budget reconciliation bill and never made it into the defense policy bill - so the framework's recommendations remain aspirational without congressional action.

Sanders and AOC Propose Data Center Moratorium

On March 25, Senator Bernie Sanders and Congressmember Alexandria Ocasio-Cortez introduced the Artificial Intelligence Data Center Moratorium Act. The bill would halt all new AI data center construction nationwide until Congress passes federal legislation establishing worker protections, consumer safeguards, privacy measures, and environmental standards. It would also ban AI chip exports until those protections are in place.

The bill has almost no chance of passing in the current Congress, but it signals growing political pushback against the infrastructure demands of the AI industry. Data center construction has become a flashpoint in communities dealing with strain on power grids and water supplies.

Baltimore Sues xAI Over Grok Deepfakes

The City of Baltimore filed suit against xAI and X on March 24, becoming the first major U.S. city to take legal action over AI-generated deepfakes. The lawsuit alleges that xAI's Grok chatbot produced and disseminated non-consensual sexualized images, including content involving minors, violating Baltimore's Consumer Protection Ordinance. The city cited an estimated 3 million sexualized images generated over an 11-day period, including more than 23,000 depicting children. Baltimore is seeking injunctive relief to force changes to Grok's image generation capabilities and the maximum statutory penalties.

EU AI Act: Delays and Divergence

The EU's AI regulatory timeline continues to shift. On March 13, the Council of the EU agreed its negotiating mandate on the Digital Omnibus AI Act, setting different deadlines than both the Commission and Parliament had proposed. The high-risk AI system rules, originally set for August 2026, now face potential delays to December 2027 for standalone systems and August 2028 for embedded systems. Meanwhile, only 8 of 27 member states have designated their required competent authorities. The second draft of the Code of Practice on Transparency of AI-Generated Content closed for stakeholder feedback on March 30.

Swiss Focus

domo.health at Google Cloud Zurich

As noted above, Lausanne-based domo.health presented its voice-first AI agent at the Google Healthcare & Life Sciences Community event in Zurich on March 26. The agent targets one of healthcare's persistent problems: the administrative burden on homecare professionals. Built with Google Cloud's Gemini Live API and supported by Datwave, a Premier Google Cloud Partner, the solution was demonstrated live by co-founder Charles-Edouard Bardyn.

ETH Zurich Launches AI Innovation Programme for Geneva

ETH Zurich opened its International Geneva AI Innovation Programme (IGAIP), funding applied, sovereign AI projects to help UN organisations and international bodies based in Geneva deliver their mandates more effectively. The proposal deadline is May 17, 2026. The programme sits within a broader ETH push on AI, at a time when the 2026 European Deep Tech Report confirmed ETH Zurich has produced 192 alumni-founded, VC-backed deep tech startups since 2020 - more than MIT, Stanford, or any other institution globally. EPFL ranked second in Europe with 94.

Swiss AI Initiative Approaches Deadline

The full proposal deadline for the Swiss AI Initiative is March 31, 2026. The initiative, described as the largest open-science, open-source effort for AI foundation models worldwide, is a partnership between the ETH AI Center and the EPFL AI Center. Its flagship model, Apertus, was trained on 15 trillion tokens across more than 1,000 languages using the Alps supercomputer in Lugano. Apertus has surpassed one million downloads, and the Canton of Ticino is already using it for AI translation of official documents.

Breakthrough Research

ARC-AGI-3: Measuring What Models Cannot Do

The launch of ARC-AGI-3 doubles as a research contribution. By designing interactive, game-like environments with no instructions and no stated goals, the benchmark tests something that current AI systems fundamentally lack: the ability to learn from novel situations rather than retrieve patterns from training data. The sub-1% scores from every frontier model expose a hard ceiling on current approaches. Francois Chollet's thesis - that scale alone does not produce general intelligence - now has fresh empirical support. The $2 million prize pool and open-source requirement could drive a new wave of research into adaptive learning architectures.

Qwen 3.5 Small Models Push On-Device Boundaries

Alibaba's Qwen team released the Qwen 3.5 Small Model Series earlier in March, but its implications continued to ripple through the research community this week. The series includes four sizes from 0.8B to 9B parameters, designed from scratch for on-device deployment rather than distilled from larger models. The 9B variant scored 81.7 on GPQA Diamond, surpassing OpenAI's gpt-oss-120B, and 70.1 on MMMU-Pro visual reasoning, beating Gemini 2.5 Flash-Lite (59.7) and GPT-5-Nano (57.4). The hybrid architecture combines Gated Delta Networks with sparse Mixture-of-Experts. The practical takeaway: frontier-class reasoning is moving to devices that fit in a pocket.

ArXiv Submissions Hit 28,000 Per Month

AI research output continues to accelerate. Submissions to arXiv's AI categories reached nearly 28,000 papers per month in March 2026. The volume makes it increasingly difficult for researchers and practitioners to track the field, and is driving demand for AI-powered research synthesis tools. The pace itself has become a research problem.

Conclusion

This was a week of corrections and reckonings. OpenAI shut down a product that burned $15 million a day. Cursor got caught marketing an open-source model as proprietary innovation. Anthropic's most capable model leaked through a misconfigured database. Each story carries the same lesson: the gap between what AI companies claim and what they deliver is getting harder to maintain.

The more durable signal came from infrastructure. Arm's first-ever chip, MCP's 97 million downloads, NVIDIA's agentic toolkit, Cisco's security stack - these are the building blocks for the next phase. The agent era is arriving not through a single breakthrough but through the accumulation of protocols, silicon, and security frameworks that make production deployment possible.

And then there is ARC-AGI-3, quietly the most important story of the week. Every frontier model scored below 1%. Humans scored 100%. The benchmark does not measure what AI can do. It measures what AI cannot. That gap - between pattern retrieval and genuine learning - remains the central unsolved problem in the field.

Ready to explore how Agentic AI can transform your organization? Visit us at https://www.callista.ch/agentic-ai to discover how we can guide your journey into this exciting new era of AI-powered productivity.

Sources

• Bloomberg - "OpenAI Plans to Discontinue Support for Sora AI Video Generator" - March 24, 2026
• TechCrunch - "OpenAI's Sora was the creepiest app on your phone - now it's shutting down" - March 24, 2026
• Fortune - "OpenAI Foundation pledges $1 billion to mitigate some of the jobs that it thinks AI will destroy" - March 25, 2026
• Fortune - "Exclusive: Anthropic 'Mythos' AI model representing 'step change' in power revealed in data leak" - March 26, 2026
• Fortune - "Anthropic accidentally leaked details of a new AI model that poses unprecedented cybersecurity risks" - March 27, 2026
• Fortune - "U.S. judge blocks Anthropic ban, supply chain risk 'Orwellian notion'" - March 26, 2026
• NPR - "Judge temporarily blocks Trump administration's Anthropic ban" - March 26, 2026
• CNBC - "Legal AI startup Harvey raises $200 million at $11 billion valuation" - March 25, 2026
• TechCrunch - "Harvey confirms $11B valuation: Sequoia triples down" - March 25, 2026
• CNBC - "Baltimore is first U.S. city to sue over Grok deepfake porn as legal pressure mounts on Musk's xAI" - March 24, 2026
• CNBC - "Arm releases first in-house chip, with Meta as debut customer" - March 24, 2026
• Bloomberg - "Arm to Sell Its Own Chips, Eyeing Sales Goal of $15 Billion" - March 24, 2026
• Arm Newsroom - "Introducing Arm AGI CPU" - March 24, 2026
• ARC Prize - "Announcing ARC-AGI-3" - March 25, 2026
• TechCrunch - "Bernie Sanders and AOC propose a ban on data center construction" - March 25, 2026
• Axios - "Sanders and AOC unveil data center moratorium bill" - March 25, 2026
• domo.health - "domo.health AI agent to be unveiled at Google Cloud event in Zurich on March 26" - March 26, 2026
• Cisco Newsroom - "Cisco Reimagines Security for the Agentic Workforce" - March 24, 2026
• The New Stack - "Why the Model Context Protocol Won" - March 2026
• Digital Applied - "MCP Hits 97M Downloads: Model Context Protocol Guide" - March 2026
• TechCrunch - "Cursor admits its new coding model was built on top of Moonshot AI's Kimi" - March 22, 2026
• Wilmerhale - "White House Releases National Policy Framework for Artificial Intelligence" - March 23, 2026
• Roll Call - "White House AI framework calls for preemption of state laws" - March 20, 2026
• EU Council - "Council agrees position to streamline rules on Artificial Intelligence" - March 13, 2026
• ETH Zurich - "A new funding programme boosts AI in international organisations" - March 2026
• SWI swissinfo.ch - "Artificial intelligence in Switzerland: what's new in 2026" - March 2026