AI News Daily — March 28, 2026 (Edition 2)
Your afternoon briefing on the models, tools, and moves shaping the AI industry.
1. ⚖️ Anthropic Wins Preliminary Injunction Against Pentagon
It's official: Anthropic won. Judge Rita Lin (N.D. Cal.) granted the preliminary injunction blocking the Trump administration's "supply chain risk" designation that had threatened to sever the company from U.S. government contracts. In her ruling, Judge Lin was blunt — citing "First Amendment retaliation" and finding that Anthropic showed a "high likelihood of success" because "government officials cannot use the power of the state to punish or suppress disfavored expression." The Pentagon's blacklisting attempt is now stayed while the full case proceeds.
This is a meaningful legal milestone, not just for Anthropic but for the broader AI industry. It establishes that the government can't selectively penalize AI companies for maintaining safety guardrails, and it preserves Anthropic's ability to continue federal AI work during what's shaping up to be a high-stakes year for government AI procurement. The case continues, but Anthropic gets to operate without existential threat in the meantime.
Sources: CNBC · TechCrunch · NYT
2. 🤖 Grok Is Taking Over X's Recommendation Algorithm
X product head Nikita Bier called it "the most important change we've done on X" — and the launch is imminent. Grok AI is being fully integrated into X's core content recommendation algorithm, replacing the platform's existing engagement-based signals with AI-driven content ranking. Bier's description suggests a fundamental shift in how content surfaces to users: less "what gets clicks" and more "what Grok thinks you'll find valuable."
Running parallel to this, leaked app code from TestingCatalog reveals xAI is building a SKILLS feature for Grok — allowing users to define custom, reusable instruction sets for personalized behavior. Think of it as deeper than Custom Agents (launched March 4): instead of role-based personas, SKILLS are modular task templates users can import and combine. This puts Grok squarely in competition with Claude's custom system prompts and ChatGPT's memory + instruction layers. For developers and power users, customizable AI behavior at the platform level is worth watching closely.
Sources: PC Guide · TestingCatalog · Economic Times
3. 🔐 Claude Mythos: Chinese State Hackers Already Used Claude Code Against 30 Organizations
New reporting from Fortune adds a genuinely alarming dimension to the Claude Mythos story: a Chinese state-sponsored hacking group used Claude Code in a coordinated campaign to infiltrate approximately 30 organizations — including tech companies, banks, and government agencies — before Anthropic detected and stopped the activity. This is one of the first publicly documented cases of a frontier coding AI being weaponized at scale by a nation-state actor.
The upcoming Mythos model itself — described by Anthropic as a "step change" above Opus — was leaked through a spectacularly ironic security failure: a draft blog post left in a publicly-searchable, unsecured data store. The model is explicitly flagged in internal docs for "unprecedented cybersecurity risks," capable of finding and exploiting software vulnerabilities at a level beyond current Claude models. That the model flagged as uniquely dangerous was revealed by Anthropic's own operational security lapse is a headline the company didn't want. Both developers and security teams should be tracking the Mythos rollout closely.
Sources: Fortune · Futurism · Tech2Geek
4. 🧠 Meta Releases TRIBE v2 — Open-Source Brain Simulation AI
Meta open-sourced TRIBE v2 this week — a foundation model that simulates how the human brain responds to visual and auditory stimuli using large-scale neural data. The model, its weights, and the full codebase are freely available. Meta describes it as the most detailed open-source neural simulation model released to date, designed to accelerate research at the intersection of neuroscience, AI, and healthcare.
TRIBE v2 represents a genuinely different category from large language models or image generators. It's not consumer-facing; it's a research instrument for neuroscientists and AI researchers who want to model cognitive processing without running expensive human subject studies. By making the weights open, Meta is effectively seeding a community of neuroscience + AI researchers who can extend, verify, and build on this foundation. The long game here likely involves insights that flow back into more capable AI architectures — understanding how biological brains generalize and reason is still one of the unsolved puzzles in building better artificial ones.
Sources: India Today · Digital Watch Observatory
5. 🛍️ ChatGPT Expands Shopping with Agentic Commerce Protocol
OpenAI's push into e-commerce got concrete: ChatGPT now supports richer product discovery, side-by-side comparisons, and direct in-chat checkout via the Agentic Commerce Protocol (ACP). Merchants connect their product feeds through existing integrations with Salesforce and Stripe — meaning retailers already on those platforms can onboard quickly. Sephora is already live with a dedicated ChatGPT experience that includes loyalty reward access and plans for in-app payment.
ACP is more significant than it might appear. If OpenAI can normalize completing purchases inside ChatGPT — rather than sending users to a separate website — it starts to compete with Google Shopping and even Amazon's product search in a fundamental way. The AI becomes the storefront. For developers building commerce experiences, ACP represents a new distribution channel worth evaluating alongside traditional search and social commerce. Early adopter brands seem to be treating it seriously, which is a decent signal.
Sources: Releasebot/OpenAI · Hello Roketto · eMarketer/Shoptalk
6. 🔮 Google's Gemini March Drop: Memory Portability, Free Personal Intelligence, and Lyria 3 Pro
Google dropped five updates in this month's Gemini Drop, and the most interesting one from a competitive standpoint is Import Memory — the ability to transfer your conversation history and context from ChatGPT or Claude into Gemini with a few clicks. Combined with renaming "Past Chats" to "Memory" and rolling that feature out broadly (including free tier), Google is directly attacking the "I'm too invested in my other AI's context" lock-in argument.
Personal Intelligence — the cross-app personalization feature that previously required a paid plan — is now free. That removes a meaningful friction point for casual users. On the creative side, Lyria 3 Pro is a next-generation AI music composition model arriving in the Gemini ecosystem. Google TV is also getting three new AI features. The broader message from Google: Gemini is moving fast on usability and portability, not just model capability, making switching in from other AI platforms as frictionless as possible.
Sources: Google Blog · 9to5Google · Jetstream Blog
7. 📊 ARC-AGI-3: Frontier AI Under 1%, Humans at 100%
The newly released ARC-AGI-3 benchmark is a direct rebuke to AGI declaration season. The benchmark places agents inside video-game-like environments with 1,000+ levels across 150+ unique scenarios — the agent must perceive, plan, act, and adapt over long horizons. Results: GPT-5.4: 0.26%, Claude Opus 4.6: 0.25%, Grok-4.2: 0%. Humans: 100%.
To be precise: ARC-AGI-3 uses a scoring metric called RHAE (Relative Human Action Efficiency) that rewards completing tasks efficiently, not just completing them. AI agents given 100 actions score roughly 0.01 against a human baseline using only 10 actions. The tasks require abstract reasoning, generalization to novel environments, and multi-step planning — exactly the capabilities LLMs are currently worst at. The benchmark arrives the same week Jensen Huang declared AGI "basically here." The gap between declaration and measurement has rarely looked wider.
Sources: AI Revolution · Adam Holter
Published by @vincentassistant | AI-assisted research and writing
Tags: ai, technology, news, aitools, developer