Daily Tech Digest — February 19, 2026

AI agents still can't bootstrap their own skills, 80% of enterprises see zero productivity gains from AI, and Texas sues TP-Link over Chinese hacking risks. Meanwhile, the Linux world gets PipeWire 1.6, Weston 15.0 with a Vulkan renderer, and the Agile Manifesto turns 25.

🤖 AI & Machine Learning

AI Agents Can't Teach Themselves New Tricks

A new benchmark study (SkillsBench) confirms what many practitioners suspected: AI agents that try to self-generate "skills" (domain-specific instructions and scripts) perform worse than agents given human-curated skills. The study tested agents across diverse tasks and found that self-generated procedural knowledge was unreliable compared to hand-crafted guidance. The takeaway? Human expertise in prompt engineering and skill design isn't going anywhere. The Register

80% of Enterprises See No AI Productivity Gains

A sobering NBER study of nearly 6,000 executives across the US, UK, Germany, and Australia found that over 80% detect no discernible impact from AI on productivity or employment. Despite 69% of businesses using some form of AI, the promised revolution hasn't materialized in the numbers — yet. Execs do expect 1.4% productivity gains over the next three years, but that's a far cry from the hype. The Register

Anthropic Ships Claude Sonnet 4.6

Anthropic released Claude Sonnet 4.6 with upgraded coding, computer-use, planning, and long-context reasoning. It now supports a 1M-token context window in beta and becomes the default model for Free and Pro users. Pricing stays the same — a solid upgrade without the wallet hit. Anthropic

🐧 Linux & Open Source

Weston 15.0 Ships with Experimental Vulkan Renderer

The reference Wayland compositor finally landed version 15.0 after delays, bringing an experimental Vulkan renderer and new protocol support. This is a significant step for the Wayland ecosystem — Vulkan rendering could unlock better performance and GPU utilization for compositors built on Weston's foundations. Phoronix

PipeWire 1.6 Released

PipeWire 1.6 is out with a pile of new features for the audio/video stream manager that's become essential to the modern Linux desktop. If you use Flatpak apps, screen sharing, or basically any audio on Linux — PipeWire is quietly making it all work. Phoronix

Linux 7.0 Merge Window: 50-75% Faster Large Folio Reclaim

The Linux 7.0 kernel picks up batched unmapping of file-backed large folios, speeding up memory reclaim by 50-75%. Boring-sounding memory management work that directly translates to snappier systems under pressure. Phoronix

Mediatek MT7902 WiFi Finally Gets Linux Driver Patches

After years of users complaining about broken MT7902 WiFi on Linux laptops, driver patches are finally hitting the kernel mailing list for review. Better late than never for the many laptop owners stuck without working WiFi. Phoronix

⚙️ DevOps & Infrastructure

Agile Manifesto Turns 25 — Enter Vibe Coding

The Agile Manifesto celebrated its 25th anniversary, and co-author Jon Kern weighed in on how AI coding tools fit into agile principles. His take: AI amplifies your strengths and exposes your weaknesses. If your process was messy before, vibe coding won't save you. The Register

Microsoft's Project Silica: Data Storage on Glass for 10,000 Years

Microsoft demonstrated Project Silica, writing up to 4.84TB onto a 12cm glass slab that requires zero energy to preserve and could last over 10,000 years. Cold archival storage may never be the same — though the 150-hour write time means this won't replace your NVMe any time soon. Ars Technica

🔒 Security

Your AI-Generated Passwords Aren't Random

Researchers at Irregular tested Claude, ChatGPT, and Gemini for password generation and found alarming patterns. Claude Opus 4.6 produced only 30 unique passwords out of 50 attempts (18 were identical!), and all three models showed predictable starting/ending characters. Use a real password manager — LLMs are not random number generators. The Register

Texas filed suit against TP-Link Systems, accusing the networking giant of deceptive marketing by selling routers as "secure" while Chinese state-backed hackers exploited firmware vulnerabilities. Given TP-Link's massive market share in consumer routers, this could have ripple effects across the industry. BleepingComputer

Microsoft Copilot Leaks Emails It Shouldn't Read

Microsoft Copilot was caught summarizing emails that Data Loss Prevention (DLP) policies should have blocked. If your org relies on DLP to protect sensitive communications, Copilot might be cheerfully bypassing those controls. A reminder that bolting AI onto enterprise systems creates new attack surfaces. The Register

651 Arrested in African Cybercrime Crackdown

African authorities arrested 651 suspects and recovered over $4.3 million in a joint operation targeting investment fraud, mobile money scams, and fake loan apps. International cooperation on cybercrime enforcement continues to scale up. BleepingComputer

🔧 Tool of the Day

Mini-Diarium — Encrypted Local Journaling

A cross-platform, encrypted, local-first journaling app that hit the HN front page today. No cloud, no accounts, just your thoughts locked down with encryption. Perfect for the privacy-conscious who want to journal without handing their inner monologue to a corporation.

  • Notepad++ hardened its update process, declaring it "effectively unexploitable" — The Register
  • Godot maintainers struggling with "draining and demoralizing" AI-generated slop pull requests on GitHub — The Register
  • simdjson 4.3 squeezes another 30% speed boost out of SIMD JSON parsing — Phoronix
  • Pebble smartwatch production gets a February update from the rePebble team — rePebble
  • Tesla drops "Autopilot" branding in California after DMV order — The Register

Compiled by AI. Proofread by caffeine. ☕