AI Dev News Digest: March 13th, 2026

Here's this week's AI dev news rounded up just for you, Enjoy!

Anthropic vs. the Pentagon

• Anthropic sued the Trump administration on Monday after being designated a "supply chain risk." The label, typically reserved for foreign adversaries, came after Anthropic refused to give the military unrestricted access to Claude for applications like mass surveillance and autonomous weapons. Defense contractors now have to certify they don't use Anthropic's models. The company says the move could cost it billions in 2026 revenue. (CNBC)
• Microsoft, Google, and Amazon all confirmed Claude stays available for non-defense customers. Nearly 40 employees from rival AI firms (including Google and OpenAI) filed a legal brief in support. (TechCrunch)
• The backlash made Claude a household name. Consumer downloads surged, paid subscriptions spiked, and the Washington Post reported Americans who'd never heard of Anthropic were racing to try it. (Washington Post)

Who Reviews the AI's Code?

Both Anthropic and OpenAI shipped code review tools this week. They landed the same week a security report found 87% of AI-generated pull requests contain at least one vulnerability. The timing wasn't coordinated, but the message is the same: AI writes code fast; nobody's checking it fast enough.

• Anthropic launches Claude Code Review. Multiple agents review each PR in parallel, cross-check findings to filter false positives, and rank issues by severity. Average review: ~20 minutes, $15–$25 per PR. Internally, Anthropic went from 16% of PRs getting substantive comments to 54%. Research preview for Team and Enterprise. (Anthropic)
• OpenAI launches Codex Security. An AppSec agent that scans repos commit-by-commit, builds a project-specific threat model, and validates findings in sandboxes before surfacing them. In 30 days of testing it scanned 1.2M commits and flagged 11,000+ high-severity issues. Evolved from an internal project called Aardvark. Free for 30 days. (OpenAI)
• OpenAI acquires Promptfoo. The AI security testing startup (used by 25%+ of the Fortune 500) will fold into OpenAI Frontier. Open-source CLI for detecting prompt injections, data leaks, and tool misuse. Previously raised $23M at an $86M valuation. (OpenAI)
• DryRun Security: 87% of AI coding agent PRs contain vulnerabilities. They tested Claude Code, Codex, and Gemini building two full apps from scratch. 143 security issues across 38 scans. Claude had the most unresolved high-severity flaws; Codex finished with the fewest. Four authentication weaknesses appeared in every final codebase. No agent produced a fully secure application. (Help Net Security)

Developer Tools

• Claude Code gets /btw. A new command that lets you have side conversations while Claude is mid-task. Ask a question about the code it's refactoring without interrupting the refactor. Also in this update: tool search works with third-party API gateways, /plan accepts a description argument, w in /copy writes directly to a file (useful over SSH), and lsof/pgrep/ss were added to the bash auto-approval allowlist. (Releasebot)
• ChatGPT adds interactive visual explanations for math and science. Ask about one of 70+ concepts and ChatGPT generates a module with adjustable sliders that update equations and graphs in real time. Available to all logged-in users, including free tier. 140 million people already use ChatGPT weekly for math and science. (OpenAI)
• Uber open-sources uSpec, an AI agent that generates Figma design specs via MCP. It connects Cursor to Figma through the open-source Figma Console MCP, crawls component structures, and generates finished spec pages (anatomy, color tokens, accessibility, screen reader mappings) in minutes. Runs locally. Uber says 31% of its code is now AI-authored and 11% of PRs are agent-opened. (Uber Blog)

Google's Big Week

• Gemini gets deeper integration across Docs, Sheets, Slides, and Drive. The Sheets update is the standout: describe a spreadsheet and Gemini builds the structure, fills data, and pulls live info from Google Search. In Docs, "Help me create" synthesizes content from Drive, Gmail, and chats. Drive gets "Ask Gemini" for natural language search with cited AI Overviews. Rolling out in beta to AI Ultra and Pro subscribers. (Google Blog)
• Google Maps gets "Ask Maps" and Immersive Navigation. Ask Maps is a conversational Gemini layer over 300M+ places and community reviews. Ask things like "is there a public tennis court with lights on tonight?" and get a custom map with directions. Immersive Navigation renders a 3D driving view with highlighted lanes, crosswalks, and stop signs, powered by Gemini analyzing Street View and aerial imagery. Google calls it the biggest navigation update in a decade. US and India on iOS and Android. (Google Blog)
• Google closes its $32B Wiz acquisition. Google's largest acquisition ever, a year after it was announced. Wiz ($1B+ ARR in 2025) joins Google Cloud but keeps its brand and will continue supporting AWS, Azure, and Oracle. Google is building a unified AI-powered cybersecurity platform combining Wiz's cloud security with its own threat intelligence. (TechCrunch)

Funding & Business

• Atlassian cuts 1,600 jobs; Adobe CEO steps down. Two SaaSpocalypse stories in one week. Atlassian is laying off 10% of its workforce (900+ engineering roles) to redirect to AI and enterprise sales, while also replacing its CTO with two AI-focused executives. Its stock has lost more than half its value in 2026. Separately, Adobe CEO Shantanu Narayen announced he's stepping down after 18 years, the same day he reported Q1 earnings. Adobe is down 23% YTD and 60%+ off its 2021 peak. (TechCrunch, CNBC)
• Grammarly sued for impersonating writers with AI "Expert Review." The feature used names and likenesses of Stephen King, Kara Swisher, Julia Angwin, and hundreds of others as AI personas, without consent. Angwin filed a class-action lawsuit and called the output a "slopperganger" because the AI feedback actually made writing worse. Superhuman pulled the feature and apologized but says it will fight the legal claims. (TechCrunch)
• Amazon wins court order blocking Perplexity's AI shopping agent. A judge temporarily blocked Perplexity's Comet from making purchases on Amazon, citing "strong evidence" of unauthorized access. Perplexity has seven days to appeal. The case could set precedent for whether platforms can refuse AI agents even when users have authorized them. (CNBC)
• Meta acquires Moltbook, the AI agent social network. The Reddit-like platform for OpenClaw agents went viral in January. Founders join Meta Superintelligence Labs. Meta wants the identity registry that ties each agent to a human owner. (Axios)

Security

• Google Cloud Threat Horizons: stolen npm token led to full AWS takeover in 72 hours. The H1 2026 report details how UNC6426 pivoted from a single GitHub token (stolen during the August 2025 Nx supply chain attack) to full AWS admin. They abused GitHub-to-AWS OIDC trust, exfiltrated S3 data, and destroyed production infrastructure. Bigger picture: software vulnerabilities overtook weak credentials as the #1 cloud attack vector for the first time (44.5% of incidents). (Google Cloud)
• AI agent hacked McKinsey's chatbot in two hours. Security startup CodeWall pointed an autonomous offensive agent at McKinsey's Lilli platform and gained read-write access to the production database via basic SQL injection. 46.5M chat messages, 728K confidential files, and 95 writable system prompts exposed. The writable prompts meant an attacker could have silently poisoned every response to McKinsey's 40,000+ users. Patched within a day. (The Register)
• March Patch Tuesday: 83 vulnerabilities, zero actively exploited zero-days. First month without one in six months. Notable: an Excel flaw (CVE-2026-26144) that could let an attacker use Copilot Agent to exfiltrate data, and two Office RCE bugs at CVSS 8.4. (CyberScoop)
• Tycoon 2FA phishing platform dismantled. Microsoft, Europol, Cloudflare, Trend Micro, and partners took down the PhaaS platform responsible for ~62% of all phishing attempts Microsoft blocked by mid-2025. 330 domains seized. The kit sold for $120 on Telegram and used adversary-in-the-middle techniques to bypass MFA in real time. (Microsoft)

Nvidia

• Nemotron 3 Super: 120B open-weights model for multi-agent systems. Only 12B parameters active during inference, delivering 5x the throughput of its predecessor. 1M-token context window to prevent goal drift in long agent workflows. Hybrid Mamba-Transformer MoE architecture, pretrained natively in 4-bit on Blackwell. Open weights, 10T+ training tokens, and RL recipes all published. Already integrated by CodeRabbit, Factory, Greptile, Perplexity, and Palantir. (NVIDIA Blog)
• GTC 2026 is next week (March 16–19). Jensen Huang teased "several new chips the world has never seen before." Expect new inference hardware, an open-model panel (Cursor, Langchain, Mistral, Ai2), and the likely formal launch of NemoClaw, Nvidia's open-source AI agent platform for enterprise. (Nvidia GTC, HotHardware, San Jose Today)

Weekend Reading

• Sam Altman muses on whether AGI should be a government project. In a weekend AMA on X, he discussed OpenAI's Pentagon contract, admitted the deal "was definitely rushed, and the optics don't look good," and speculated about the government nationalizing AI companies. OpenAI's robotics leader Caitlin Kalinowski resigned over the deal. (The New Stack)
• The 8 Levels of Agentic Engineering. Bassim Eledath maps the progression from tab completion to fully autonomous background agents. The key insight: if you're running background agents but your colleague still reviews manually, that bottleneck kills your throughput. (Bassim Eledath)
• "I built a programming language using Claude Code." Ankur Sethi built Cutlet, a complete dynamic language with a REPL, garbage collector, and prototypal inheritance, in four weeks. He never read the generated code. He built guardrails (tests, specs) and let Claude write everything. (Ankur Sethi)
• No, it doesn't cost Anthropic $5k per Claude Code user. Martin Alderson debunks the Forbes claim. The $5k figure confuses retail API prices with actual compute costs. Real compute cost for the heaviest users is closer to $500/month. (Martin Alderson)
• The kernel gets ready for Rust in drivers. LWN covers the push to bring Rust into the Linux kernel's driver subsystem and the friction between Rust and C maintainers. (LWN)
• Microsoft Copilot update hijacks default browser links. A recent update reportedly forces links to open in Edge regardless of your default browser. (Reclaim The Net)
• Kapwing is paying artists royalties for AI-generated art. What they've learned building a royalty system that compensates artists whose work influenced AI outputs. (Kapwing)
• We're training students to write worse to prove they're not robots. TechDirt on the perverse loop where AI detection tools push students toward deliberately awkward writing, which then pushes them toward using more AI. (TechDirt)
• How Is Felix Today? A personal site tracking one developer's daily status, mood, and tools. (howisfelix.today)
• I'm building agents that run while I sleep. Claude Code Camp on setting up autonomous agents for overnight coding sessions. (Claude Code Camp)

Weekend Watch

• Agent Skills or MCP? (Confluent Developer, Tim Berglund). Tim recaps what's new in the MCP spec (streamable HTTP, OAuth 2.1, the quiet decline of the Resource API), breaks down what skills are, and explains where they overlap with MCP. The short answer: local coding agents can lean on skills, but agentic microservices still need MCP. (YouTube)

---

Read next: Inside Look at Using Claude Code Remote Control

Recommended

Recommended

Inside Look at Using Claude Code Remote Control

Anthropic's Remote Control, currently in research preview, adds a third option for Claude Code users who need to step away mid-session. Your session keeps running on your machine while your phone becomes a window into it…

Read next