Sec & AI News — 20 February 2026

20 February 20268 min readBy Nathan House

Five Frontier Models Shipped in One Week

Grok 4.20 Runs Four Agents That Argue First

OpenAI Deep Research Jumps to GPT-5.2

Gemini Now Makes Music. Thirty Seconds of It.

Seedance 2.0 Has Hollywood Rattled

Pentagon Wants Out of Anthropic Over Guardrails

China's Autonomous War Machines Get a Deadline

Jobs, Money, and the Productivity That Never Showed

Meta Patented an AI That Posts After You Die

Five Frontier Models Shipped in One Week

Get every new Sec & AI News issue Straight to your inbox. No spam.

🟣 Five Frontier Models Shipped in One Week

Gemini 3.1 Pro, Claude Sonnet 4.6, Grok 4.20, GPT-5.3 Codex, Qwen 3.5. All in February. Google claims leading scores on 13 of 16 benchmarks and 77.1% on ARC-AGI-2. Sonnet 4.6 is the new Claude.ai default and tops the GDPval-AA Elo board at Sonnet pricing. Nobody slept.

🔵 Grok 4.20 Runs Four Agents That Argue First

xAI shipped something weird. Four specialised agents — Grok, Harper, Benjamin, Lucas — debate every complex query before answering. Baked into inference, not a workflow. In Alpha Arena, trading real capital, Grok 4.20 was the only profitable model. And this is just the 500B "small" variant.

🔍 OpenAI Deep Research Jumps to GPT-5.2

Deep Research got bumped from o3 to GPT-5.2. Now you can prioritise specific sites in the search. Available to all paid ChatGPT users. Tell it a topic, it browses, comes back with a 4,000-word cited report. Whether the citations are real is your problem.

🎵 Gemini Now Makes Music. Thirty Seconds of It.

Lyria 3, Google DeepMind's music model, landed in beta inside the Gemini app. Describe a vibe, upload a photo, get a track with AI lyrics and cover art. Thirty-second clips. Not replacing your Spotify. Replacing your group chat with personalised afrobeat about your dog.

🎬 Seedance 2.0 Has Hollywood Rattled

ByteDance dropped Seedance 2.0. The clips are realistic enough to make studios nervous. SAG-AFTRA is already siding with the studios against it. AI video just went from "neat demos" to "intellectual property crisis."

🔴 Pentagon Wants Out of Anthropic Over Guardrails

The US military used Claude during the Maduro raid in Venezuela. Then Anthropic's safety restrictions kicked in. The Pentagon is reportedly shopping for alternatives because Claude won't do what they want. The company that spent years building guardrails just learned its biggest government customer calls them a bug.

🤖 China's Autonomous War Machines Get a Deadline

China's autonomous military robots are reportedly two years from the battlefield. Not a concept video. An actual timeline. Meanwhile Amodei declined to hold Altman's hand for a summit photo, and the clip went viral. Former colleagues. Current competitors. Awkward doesn't cover it.

💰 Jobs, Money, and the Productivity That Never Showed

Microsoft pledged $50 billion for the "growing AI divide" the same week its AI chief said all white-collar work could be automated in 18 months. IBM plans to triple entry-level US hiring — same jobs, recast for the AI era. And 80% of companies report no measurable productivity gains. The Solow Paradox rides again. ROI still loading.

💀 Meta Patented an AI That Posts After You Die

Meta was granted a patent for an AI that keeps generating posts from your account after death or prolonged absence. Your digital ghost, LLM-powered, trained on your posting history. Every tech dystopia novel just became a patent filing.