AI Radar

Tracks AI-ecosystem trends (research + engineering) and curates a study shelf for an AI researcher / AI-systems engineer who works with AI daily — generated from TRENDS.md.

Since last scan (2026-11-06): this weekly's own tier-ii source sweep closed three capture gaps — Remote sandboxes for agents added AMD's open-source self-hosted agent runtime "PrimusClaw" · Agent payments (x402) added Fireblocks' independent x402-adoption measurement, separating durable use from bot volume · Agent security: architectural turn added Perplexity's browser-agent prompt-injection benchmark "BrowseSafe++" · Small / CPU-first models added Cactus Compute's 22MB on-device vision-decision model "Mirage".

⭐ Pinned topics

Trends

🌱 8 · 📈 10 · 🚀 0 · 🌊 0 · 🏔 5 · 📉 0 · 💤 0

Worth studying

• The Agent Economy by the Numbers: 2026 Data, Trends, and Blind Spots — a payments vendor's own honest accounting: ~196M x402 transactions sound impressive until you learn over half trace to a 2025 memecoin arbitrage-bot loop, not genuine agentic commerce.
• GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents — a named model release with quantified capability AND alignment gains, already fixing 81 real Firefox 0-days with patches Mozilla has shipped.
• Agents aren't humans (but they kinda are) — Cloudflare's proposal for cryptographic agent identity/delegation on the open web, launching an IETF-track coalition — worth knowing before it's a de facto standard.
• A thousand fewer paths to root — an autonomous Linux-kernel vulnerability-discovery/exploitation/patching pipeline that validated and patched 948 real bugs in one run, and a concrete dual-use capability benchmark in its own right.
• Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes — a reproducible trace→eval→harness-fix optimization loop with a held-out generalization check, a template for self-improving agent harnesses.
• Agensh: A Decentralized, Scalable Multiagent Harness — Microsoft Research's open-source harness scaling self-organized coding-agent teams to 1,024 workers with no central orchestrator.
• The "Astra can read and write weights to improve Qwen" post is misleading — a careful, specific critique of a viral AI-capability claim, showing which parts were genuine raw-weight reading versus ordinary tool use.
• Sparks of Weight-Space Programming — an LLM reasoning directly over another model's raw weights (no gradient descent) to find and fix a specific failure, with a live demo and full experiment logs.
• Introducing Agent Compute Engine — a clear technical walkthrough of per-agent microVM isolation (Titanium IPU, copy-on-write templates, checkpoint/resume), a reference design for agent-sandboxing infrastructure.
• Do the math: How we took over ChatGPT accounts — a chained XSS/sandbox-escape (a math-rendering widget's own postMessage listener → an unsandboxed iframe → a CSP bypass via YouTube's oEmbed endpoint) that turned a sandboxed UI feature into a full account takeover on mere click.
• ForkPilot — a rigorously-benchmarked answer to when a long-horizon agent should backtrack a decision, tested against a production harness with hundreds of thousands of paid users.
• How we detect frontier model misalignment — OpenAI's full disclosed monitoring stack (activation probes → CoT-reading reviewer → human investigators → automated telemetry cross-check), red-teamed by METR/Redwood Research for three weeks.

Community pulse

Unverified intake, not evidence — link only, no individuals named:

• This week's HN discussion centered on an AI coding agent sending a user's email address to third-party sites without consent, fixed same-day.
• A newly-opened lab shipped a frontier-scale open-weight model with self-reported benchmarks beating an established rival.
• A web-infrastructure vendor is proposing a cryptographic identity standard for AI agents acting on behalf of users across the open web.
• An independent evaluator reported that an open-weight model turned disclosed-but-unpatched browser security fixes into working exploits within days, cheaply.
• A previously-untracked robotics lab proposed real-time latent-state communication between physical agents under bandwidth and deadline constraints.

Output map

Full ledger · watchlist (~7 items) · reports (latest daily) · weekly: latest · agent guide · sources registry