Radar Trends to Watch: January 2026

Loukides — Radar Trends to Watch: January 2026

The first installment in the wiki of Mike Loukides’s monthly Radar Trends to Watch digest at O’Reilly Radar — a curated, six-section roundup of the prior month’s tech news, with editorial weighting toward the threads Loukides expects to compound. Published 6 January 2026, covering events from the December 2025 holiday window.

Opening framing notes that December was truncated by O’Reilly’s holiday break, but two foundational reads are recommended: Andrej Karpathy’s LLM year review (the upstream conceptual spec for this wiki) and the Resonant Computing Manifesto. OpenAI’s year-end push to release GPT-5.2 and to secure Disney’s character-licensing deal for Sora is named as the headline story.

TL;DR — the month’s standout threads

• Agent Skills become an open spec. Anthropic opened the Claude Skills spec; OpenAI quietly adopted them. The Skills-as-portable-package primitive is becoming cross-vendor.
• GPT-5.2 released in three versions targeting professional knowledge workers — the year-end model push that frames January’s framing.
• Gemini 3 Flash named as Google’s “final” model combining reasoning, speed, and economy.
• Rust enters the Linux kernel without the experimental flag; Tor Project ships Arti (its Rust rewrite) as production-ready. Ruby 4.0 released on Christmas (Matz tradition).
• “Cognitive Architect” role surfaces — developer decomposes problems and emphasises higher-order thinking. Adjacent to the wiki’s agentic-engineering vocabulary.
• PARK stack (PyTorch + AI + Ray + Kubernetes) named as the shape of open-source AI development infrastructure.
• Prompt injection consolidates as a first-class security threat. Chrome ships a “User Alignment Critic” that monitors Gemini actions for indirect prompt injection (the wiki’s first browser-side prompt-injection countermeasure source). MITRE’s Top 25 weaknesses 2025 retains XSS / SQL injection / CSRF on the list — normalization of deviance in AI surfaces as a security framing.
• Poetry as LLM jailbreak. Writing prompts as verse evades model defenses — a new attack pattern Loukides flags.

Sectional summary

AI (14 items). Function-calling adaptations (FunctionGemma — Google’s function-calling 270M Gemma 3); Agent Skills spec opened by Anthropic; Gemini 3 Flash; GPT-5.2; Disney–OpenAI Sora licensing deal; Nemotron 3 Nano (NVIDIA, 30B MoE, fully open source); Google’s Titans architecture (efficient memory handling); Mistral 3 + three smaller variants with open weights; DeepSeek V3.2 (sparse attention + RL + task synthesis); FLUX.2 (Black Forest Labs vision model approaching Google Nano Banana quality).

Programming (9 items). Ruby 4.0 Christmas release; Tor Project’s Rust rewrite (Arti); Cognitive Architect role; Rust in Linux kernel out of experimental status; PARK stack (PyTorch / AI / Ray / Kubernetes); Go vs Rust vs Zig technical comparison; Stack Overflow AI Assist (conversational search for Stack Exchange); DocumentDB open source (MIT-licensed Mongo-/Postgres-hybrid); Christina Wodtke argues user experience as moat, not model.

Security (7 items). SantaStealer malware (rebranded BluelineStealer; browsers + wallets + services); MITRE Top 25; normalization of deviance in AI (false safety sense around prompt injection); AI-edited image consequences (a bridge-collapse image cancelled train services); virtual-kidnapping scams (doctored social media used as ransom “proof”); poetry as LLM jailbreak; GrayNoise IP-check tool (free botnet IP scanner).

Web (5 items). Chrome User Alignment Critic (monitors Gemini for indirect prompt injection); Android XR smart glasses (Google prototypes monocular/binocular Gemini-backed frames); Lightpanda (browser engineered for crawlers / automated apps); Nook Browser (open-source privacy-preserving fast browser for humans); Brave AI-Assisted Mode (Leo for privacy-preserving agentic multistep tasks).

Hardware (2 items). Arduino vs Adafruit licensing controversy; NYU researchers on democratising microchip design.

Operations (1 item). Cloudflare outage analysis — detailed examination of November’s internet-wide disruption.

Cross-positioning with the wiki

• Sub-monthly granular feed for Signals for 2026 — Baron’s annual outlook crystallises three days later; the January digest is the underlying news log. MCP, Skills, agentic dev, cognitive architect, prompt injection appear in both.
• First “User Alignment Critic” mention in the wiki — a structural countermeasure to prompt injection at the browser/agent boundary. Will surface again across multiple monthly digests as agent-side defences mature.

Named entities (this ingest)

• Mike Loukides — O’Reilly Radar editor; recurring author of this monthly digest. First wiki mention; promoted to entity in this batch (5 source pages in batch).
• O’Reilly Media — Publisher. Promoted to entity in this batch.

Plus dangling first-mentions of: GPT-5.2, Gemini 3 Flash, Mistral 3, FLUX.2, Nemotron 3 Nano, DeepSeek V3.2, Titans architecture, PARK stack, Christina Wodtke, MITRE.

Source-quality notes

• Genre: monthly editorial-curatorial digest. Not peer-reviewed; not empirical. Items are pointer-summaries with embedded links to original sources; this source page captures the editorial weighting and threading rather than the underlying news.
• Confidence: 0.65. Per Lifecycle: single secondary-summary source, +0.05 for O’Reilly editorial bar; the underlying claims trace to the embedded links rather than to Loukides’ own analysis. Citations from this page should reference the underlying primary source where available; the digest page itself is best used as a thread index into the month’s trend space.