Sec & AI News — 6 March 2026
🚀 OpenAI Ships GPT-5.3 and GPT-5.4 Back to Back
Two drops in one week. GPT-5.3 Instant got fine-tuned to moralize less — fewer refusals, fewer caveats, benchmarks unchanged. GPT-5.4 is the one that matters: native computer use, a token-slashing tool search, a 1M context window, and a nudge better at coding. Chatbot users get marginal gains. Agent builders get the whole point.
🧩 Google Ships Flash Lite, Cinematic NotebookLM, and Search Canvas
Google went wide this week. Gemini 3.1 Flash Lite trades benchmarks for millisecond speed and cheap tokens — the backbone nobody tweets about. NotebookLM now spins your notes into real animated explainers via VO3 and Gemini 3, gated behind the $250/month Ultra plan. And AI Mode in Search gained a Canvas panel that renders what you ask for live. Same idea as Claude Artifacts. Google's late, but Google has distribution.
- Gemini 3.1 Flash Lite — Google Blog
- NotebookLM Cinematic Video Overviews — Google Blog
- Canvas in AI Mode — Google Blog
💥 Pentagon vs Anthropic: Red Lines Meet Supply Chain Risk
Anthropic held two lines: no domestic surveillance, no autonomous weapons without a human in the loop. The Pentagon said drop them. Anthropic refused and got designated a supply chain risk — blacklisted from federal contracts. OpenAI arrived the same day with identical red lines, plus a third, and got approved. A leaked Anthropic memo called that deal "safety theater," Dario walked it back, and Altman backed Anthropic publicly while telling staff they have "no say" over Pentagon decisions. Anthropic says the designation isn't legally sound and will fight it in court.
- Anthropic's Statement on Supply Chain Designation
- OpenAI's Agreement with the Department of War
- Pentagon Designates Anthropic a Supply Chain Risk — Bloomberg
- Altman Tells Staff OpenAI Has No Say Over Pentagon Decisions — Bloomberg
- Altman Defends Pentagon Work, Calls Backlash "Really Painful" — WSJ
- OpenAI Will Amend Deal to Prevent Mass Surveillance — Engadget
📱 ChatGPT Exodus: Uninstalls Up 295%, Claude Hits Number One
Users voted with their thumbs. ChatGPT uninstalls surged 295% over a weekend and Claude jumped to #1 in the App Store. Anthropic shipped a memory import tool — bring your ChatGPT data, switch for free — and put memory on the free plan too. Enterprise API spend flipped in Anthropic's favor, with revenue nearing a $20B run rate. The Pentagon drama did what no competitor could.
- ChatGPT Uninstalls Surged 295% After DoD Deal — TechCrunch
- Users Ditching ChatGPT for Claude — TechCrunch
- Anthropic Nears $20B Revenue Run Rate — Bloomberg
- Claude Memory Now Free for All Users — Engadget
🐲 Open Weights Everywhere: Qwen 3.5, Grok 4.20, Phi-4 Vision
The open-and-iterative crowd kept shipping. Alibaba's Qwen 3.5 comes in four sizes from 800M to 9B — the smallest runs on an iPhone offline, the 9B reportedly beats models ten times larger. Microsoft's Phi-4 Reasoning Vision is a 15B open-weight multimodal that reasons across math, science, and UI. And Grok 4.20 Beta 2 landed with better instruction following and fewer hallucinations. Not revolutionary. Still shipping.
- Qwen 3.5 — Official Blog
- Alibaba's Qwen3.5-9B Beats Much Larger Models — VentureBeat
- Phi-4 Reasoning Vision — Microsoft Research
- Grok 4 — xAI
- Introducing the Codex App — OpenAI
- OpenAI Codex App Comes to Windows — WinBuzzer
👓 Surveillance Hardware: Meta Glasses Leak, and a Device to Shut Them Up
Two sides of the same anxiety. An investigation found Meta AI glasses stream footage — credit cards, bathrooms, everything — to human annotators in Africa when privacy settings aren't set right, with lawsuits in New Jersey and California and the UK's ICO now investigating. Meta says opt-in; workers say "we see everything." Meanwhile Deveillance launched Spectre I, a $1,199 gadget that scrambles audio so nearby devices can't record within two metres. In a world that always listens, someone's selling silence.
- UK Data Watchdog Investigating Meta Glasses — BBC
- Meta Sued Over AI Smartglasses Privacy — TechCrunch
- Swedish Investigation: Workers Say "We See Everything" — SVD
- Deveillance Spectre I — Official Site
- This Device Stops AI Gadgets From Listening — Digital Trends
🟢 NVIDIA GTC 2026: March 16-19, San Jose
Jensen Huang keynotes March 16. New chip expected. Sessions on AI, robotics, and everything NVIDIA wants you to buy. Virtual attendance is free.