Volume 1, No. 26 Thursday, March 26, 2026 Daily Edition

The AI Dispatch

“All the AI News That’s Fit to Compile”


Exclusive Leak

Anthropic’s Next Frontier Model ‘Mythos’ Exposed in Data Breach — Company Calls It ‘A Step Change’

Nearly 3,000 unpublished assets, including a draft blog post for an unreleased model codenamed “Capybara,” were discoverable in an unsecured data cache — triggering Anthropic’s first public acknowledgment that Mythos exists.

Anthropic finds itself disclosing its most capable model earlier than planned after Fortune reported that nearly 3,000 unpublished internal assets — including a complete draft blog post for an unreleased frontier model — were accessible in an unsecured data store. The company acknowledged the existence of Claude Mythos, also referred to internally as “Capybara,” confirming that training has been completed and that early access testing is currently underway with select customers.

An Anthropic spokesperson described Mythos as representing “a step change” in AI performance and “the most capable we’ve built to date.” Leaked documents indicate the model significantly outperforms its predecessors in software coding, academic reasoning, and — most consequentially — cybersecurity tasks. The company acknowledged in internal materials that Mythos is “currently far ahead of any other AI model in cyber capabilities” and expressed concern that it could enable large-scale cyberattacks if deployed without care.

Anthropic says it is “being deliberate about how we release it” precisely because of these cybersecurity risks. The leaked materials also surfaced plans for an exclusive CEO summit in Europe at which unreleased Claude capabilities were to be demonstrated to business leaders. No public launch date has been announced. The incident highlights a deepening tension in frontier AI development: as models grow more capable, the gap between completion and public release widens — and unsecured data stores become a liability the entire industry must address.

First Amendment Win

Federal Judge Blocks Pentagon’s Anthropic Blacklist, Calls It ‘Classic First Amendment Retaliation’

U.S. District Judge Rita Lin granted a preliminary injunction barring the Defense Department from enforcing its “supply chain risk” designation against Anthropic, ruling the ban was retaliation for public speech.

In a 43-page ruling, U.S. District Judge Rita Lin granted Anthropic a preliminary injunction blocking the Pentagon from enforcing the “supply chain risk” designation that Defense Secretary Pete Hegseth applied to the company in late February. The designation — historically reserved for foreign adversaries — made Anthropic the first American company publicly labeled a potential national security threat under the statute. Federal agencies were prohibited from using Claude following the designation.

The dispute arose from contract negotiations that stalled after Anthropic signed a $200 million agreement with the Pentagon in July and then refused to grant the DoD “unfettered access” to its models for all lawful purposes, insisting on guardrails against use in fully autonomous weapons and domestic mass surveillance. When talks broke down, the Trump administration escalated to the supply chain designation.

Judge Lin was unsparing in her assessment: “Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government.” She added: “Punishing Anthropic for bringing public scrutiny to the government’s contracting position is classic illegal First Amendment retaliation.” The ruling is a preliminary injunction, not a final judgment; the underlying lawsuit will continue. OpenAI and Google employees filed amicus briefs supporting Anthropic earlier in the proceedings.

Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government. U.S. District Judge Rita Lin — Preliminary Injunction Ruling, March 26, 2026

Industry & Market

Market Shift

Claude Now Wins 70% of First-Time Business AI Matchups Against OpenAI, Ramp Data Shows

The March 2026 Ramp AI Index documents the sharpest competitive shift yet recorded in enterprise AI adoption. Among businesses purchasing AI services for the first time, Claude wins approximately 70% of head-to-head matchups against OpenAI. A year ago, one in 25 companies on the Ramp platform paid for Claude; today it is nearly one in four, representing a 4.9% month-over-month gain — Anthropic’s largest single-month jump since tracking began.

OpenAI simultaneously recorded its worst monthly performance: a 1.5% decline in business subscriptions, the steepest drop for any AI model company in Ramp’s dataset. Yet OpenAI retains overall market share leadership at 34.4% versus Claude’s 24.4%, underscoring that the race is ongoing. The data suggests Claude’s wins are concentrated in two areas: developer API spend (where Anthropic leads at roughly 80%) and first-impression enterprise evaluations, where Claude’s coding and reasoning capabilities are proving decisive.

xAI

Musk’s xAI ‘Doubling Down’ on AI Video as OpenAI Cedes the Field

With OpenAI having shuttered Sora, Elon Musk announced on March 25 that xAI intends to aggressively expand Grok Imagine’s video generation capabilities, calling the next release “epic.” The move represents an opportunistic land-grab in the AI video market at a moment when the dominant incumbent has stepped aside.

Regulation

Dutch Court Orders xAI to Stop Generating Nonconsensual Nude Images Under €100K Daily Fine

An Amsterdam District Court ruled on March 26 that xAI’s Grok is barred from generating or distributing sexual imagery of people who have not given explicit permission, imposing fines of 100,000 euros per day for noncompliance. The ruling represents one of the first judicial orders directly targeting an AI model’s image generation behavior in the EU, and will test whether a Dutch court can effectively compel an American AI company to modify its product internationally.

Science & Research

Milestone

The AI Scientist Makes History: First Fully AI-Generated Paper Passes Peer Review, Published in Nature

Sakana AI’s AI Scientist-v2, a fully autonomous research agent, achieved what its creators call a landmark milestone: an entirely AI-generated paper was submitted to the ICLR 2025 ICBINB workshop, scored an average of 6.33 by human reviewers — above the acceptance threshold and higher than 55% of human-authored submissions — and was accepted. The researchers obtained permission and preemptively withdrew the accepted paper. The system autonomously generates research ideas, searches literature, designs and runs experiments via parallelized agentic tree search, and writes full LaTeX papers with figure feedback.

Version 2 marked a significant leap over v1 by removing the requirement for starter code templates, allowing the agent to investigate broadly defined topics with genuine freedom. The team also built an Automated Reviewer achieving 69% balanced accuracy — comparable to human reviewers at AI conferences. The paper describing the full system was published in Nature on March 26.

Neuroscience AI

Meta Open-Sources TRIBE v2: An AI ‘Digital Twin’ of the Human Brain That Predicts Neural Responses

Meta’s Fundamental AI Research team released TRIBE v2 (TRImodal Brain Encoder), a foundation model trained on 1,117 hours of fMRI data from 720 subjects watching movies, listening to podcasts, and viewing silent video. The model predicts how any individual brain will respond to new stimuli — including previously unseen languages — without retraining, achieving a 70-fold resolution improvement over prior systems. Meta released the model, codebase, paper, and an interactive demo to the global scientific community.

The applications range from neuroscience research (simulating neural responses without requiring new brain scans) to clinical medicine (better understanding neurological disorders like aphasia) and potential brain-computer interface development. The open-source release marks one of the most comprehensive contributions to computational neuroscience from a commercial AI lab.

Quick Dispatches

Qwen 3.5 Small Series: 9B That Beats 120B

Alibaba’s Qwen team released the Qwen 3.5 Small Model Series (0.8B to 9B parameters) for on-device and edge deployment. The flagship 9B model achieves a GPQA Diamond score of 81.7, outperforming OpenAI’s gpt-oss-120B — a model thirteen times larger. All sizes are available on Hugging Face and ModelScope. VentureBeat

MCP Crosses 97 Million Installs

Anthropic’s Model Context Protocol crossed 97 million cumulative installs in March 2026, reaching that milestone faster than most developer infrastructure protocols achieve in their first five years. The protocol now underpins the Agentic AI Foundation under the Linux Foundation, alongside OpenAI’s AGENTS.md and Block’s goose framework. ClawPod

Gemini Code Assist Goes Free for Individual Developers

Google made Gemini Code Assist fully free for individual developers, providing complete access to the IDE plugin for VS Code and JetBrains at no cost. The move directly challenges GitHub Copilot’s individual tier and intensifies the AI coding assistant price war that has already driven several competitors to reduce or eliminate subscription fees. Digital Applied

obra/superpowers Hits 113k Stars as Agentic Skills Surge

The obra/superpowers Shell framework — an agentic skills methodology for Claude Code with a marketplace of reusable skills — reached 113,500 GitHub stars this week, up from 108k reported just days ago, reflecting extraordinary developer appetite for structured agentic workflows. It is currently one of the fastest-growing repositories on GitHub. Trendshift

Copilot CLI v1.0.12 Ships With MCP and Full-Screen Model Picker

GitHub released Copilot CLI 1.0.12 on March 26, fixing MCP server startup when the working directory is the git root, enabling clipboard copy on Windows, and adding a model header that shows the active reasoning effort level. The update also adds full-screen model selection with inline reasoning effort adjustment. Releasebot

Grok 5 Anticipation Builds as xAI Eyes Late Q2 Release

As xAI doubles down on video and faces its Dutch court ruling, developer community chatter around Grok 5 is intensifying, with reports suggesting a late Q2 2026 target. The release would follow Grok 4.20’s earlier-than-expected rollout and is expected to feature substantially improved reasoning, longer context windows, and deeper integration with X platform data. NxCode

Toolbox

Developer Tool Changelog: Copilot CLI & Claude Code Ecosystem

GitHub Copilot CLI v1.0.12 (March 26)

MCP server fix — Servers now start correctly when the working directory is the git root, resolving a common failure case for monorepo users. Windows clipboard copy now works reliably. Model display header shows the active reasoning effort level at a glance. Full-screen model picker with inline reasoning effort adjustment (select model + tune effort in one step). Reading .claude/settings.json and .claude/settings.local.json as repo config sources.

Claude Code Ecosystem (Ongoing March)

The Yeachan-Heo/oh-my-claudecode open-source project (11.8k stars, TypeScript) provides teams-first multi-agent orchestration for Claude Code — currently one of the fastest-climbing repos on GitHub. Separately, the msitarzewski/agency-agents framework (62.7k stars) offers a complete AI agency with specialized agents for distinct development roles, showing strong enterprise uptake this week.

Codex CLI (No new release in last 48h)

No releases in the 48-hour window, though Claude and Codex agents remain available for Copilot Business and Pro subscribers directly within github.com, GitHub Mobile, and VS Code as of the February GA launch.

GitHub Trending

Repository Language Stars Description
obra/superpowers Shell 113.5k Agentic skills framework & software development methodology for Claude Code
anthropics/claude-code Shell 82.7k Terminal-based agentic coding assistant for faster development with natural language
hacksider/Deep-Live-Cam Python 80.3k Real-time face swap and deepfake video generation from a single image
msitarzewski/agency-agents Shell 62.7k Complete AI agency with specialized agents for distinct development roles
Yeachan-Heo/oh-my-claudecode TypeScript 11.8k Teams-first multi-agent orchestration framework for Claude Code
Vaibhavs10/insanely-fast-whisper Jupyter Notebook 11.6k Optimized speech-to-text pushing Whisper to its performance limits
mvanhorn/last30days-skill Python 8.4k AI agent skill that researches any topic across Reddit, X, YouTube, HN, and Polymarket

Source: Trendshift • Star counts as of March 26, 2026