Key Signals

• Claude Code's explosive growth is reshaping Anthropic's business and the broader AI coding market. Claude Code usage surged 300% since the Claude 4 models launched, with run-rate revenue up 5.5x. An enterprise analytics dashboard now tracks spend and code acceptance rates. Independently, credit card transaction data analyzed by Indagari shows Claude paid subscriptions more than doubled this year, with Claude Code and Cowork identified as primary subscription drivers — suggesting developer tooling is becoming a major consumer acquisition channel for AI labs. [1][2]

• AI coding tools have crossed an inflection point: they are now finding real bugs in open source at scale. Linux kernel maintainer Greg Kroah-Hartman reported that AI-generated security reports suddenly shifted from "slop" to legitimate findings across all major open source projects roughly one month ago. In his own experiments, two-thirds of AI-generated patches were correct. Google's Sashiko code review tool, now donated to the Linux Foundation, is being integrated into kernel review infrastructure — signaling that AI review is moving from experiment to standard workflow. [3]

• Nvidia's NemoClaw exposes the unsolved security problem at the heart of agentic AI coding. NemoClaw adds three security layers on top of OpenClaw — policy enforcement, privacy routing, and sandboxed execution — but analysis argues none address the fundamental tension of autonomous agents that install packages, learn skills, and spawn subagents. The proliferating "claw" ecosystem reflects growing enterprise demand, but bolt-on security may not be sufficient for production deployment. [4]

• MCP hit 97 million monthly SDK downloads, cementing its role as the standard protocol for agentic tools. The Model Context Protocol grew 4,750% in 16 months from roughly 2 million downloads at launch to 97 million in March, with over 5,800 community and enterprise servers now in the ecosystem. After OpenAI adopted MCP and Anthropic donated it to the Agentic AI Foundation under the Linux Foundation, the protocol is no longer a proprietary play — it is the industry's shared infrastructure for agent-tool communication. [1]

• Anthropic's leaked "Claude Mythos" model hints at another step change in AI coding capability. An accidental data cache exposure revealed Anthropic is testing a next-gen model codenamed "Capybara" with dramatically higher scores on coding, academic reasoning, and cybersecurity benchmarks compared to Opus 4.6. Anthropic itself warned the model is "far ahead of any other AI model in cyber capabilities" and "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." [1]

AI Coding News

• Anthropic shipped 14+ major features in March while battling five outages and an accidental model leak. A comprehensive roundup catalogues what may be the most aggressive shipping streak in AI history: Claude Code multi-agent review (54% of PRs now get substantive comments, up from 16%), Claude Code on web and mobile, Claude Dispatch for persistent agent tasks, computer use in research preview, 1M-token context at standard pricing, the Claude Marketplace, and the Claude Partner Network backed by a $100M commitment. The pace came at a cost — Anthropic experienced at least five service outages in March, including two in the final week, raising questions about whether reliability is keeping up with feature velocity. [1]

• Claude paid subscriptions are skyrocketing, driven by developer tools and the DoD controversy. Credit card transaction analysis of 28 million U.S. consumers shows Claude gaining paid subscribers at record rates, with the sharpest growth between late January media coverage of the Anthropic-DoD feud and CEO Dario Amodei's public statement on February 26. Most new subscribers are at the $20/month Pro tier. Anthropic confirmed paid subscriptions more than doubled this year, with Claude Code, Cowork, and the new Computer Use feature identified as key growth drivers. Despite this momentum, Claude remains well behind ChatGPT in total paid consumer base. [2]

• AI code review tools have gone from nuisance to necessity for open source maintainers. Greg Kroah-Hartman, speaking at KubeCon Europe, described a sudden shift roughly one month ago where AI-generated security reports went from "obviously wrong" to legitimate across every major open source project. "Something happened a month ago, and the world switched. Now we have real reports," he said. The Linux kernel is integrating Sashiko, a Google-developed AI review tool now under the Linux Foundation, into its standard workflow. Different subsystems are contributing specialized prompts — storage, graphics, networking each have tailored review skills. The immediate win is turnaround time: AI reviewers flag obvious problems before human maintainers would see the patch. The challenge is that smaller projects lack the capacity to absorb the growing flood of AI-generated reports, even when those reports are now genuinely useful. [3]

• Nvidia's NemoClaw security architecture gets a skeptical analysis from The New Stack. NemoClaw's three-layer approach — policy enforcement constraining filesystem/network access, privacy routing to control what goes to cloud vs. local models, and sandboxed execution for process isolation — is described as necessary but insufficient. The core argument: "There is an inherent inefficiency in letting self-evolving agents install packages, learn skills, and spawn subagents only to stop them at the door because you don't like what they are wearing." The article notes the real bottleneck for enterprise agentic adoption will be hiring experienced engineers to manage risk profiles, not the availability of tools. [4]

Feature Update

• Gemini CLI v0.35.3 and v0.36.0-preview.6 released with P1 sandbox network fix. Both releases cherry-pick the same commit from PR #24055 — "Enable network access in sandbox configuration" — a P1 priority bug fix that corrects the default behavior preventing sandboxed tools from making network requests. The fix was authored by @galz10 and merged on March 27, then immediately cherry-picked to both the stable (v0.35.3) and preview (v0.36.0-preview.6) release channels on March 28. This fix is notable because it touches the sandbox configuration that Gemini CLI recently shipped as part of its new macOS Seatbelt, Linux bubblewrap, and Windows sandboxing infrastructure — a core feature differentiating Gemini CLI's security posture. [5][6]