Radar Trends to Watch: May 2026

Loukides — Radar Trends to Watch: May 2026

The fifth (and most recent) installment of the monthly digest, published 5 May 2026, covering April 2026. Co-authored with Claude (first time the digest carries an explicit human-AI co-byline). Opening framing — “AI is becoming operational”: the highlight tension is between Anthropic’s restricted approach to Claude Mythos (limiting access through Project Glasswing) and OpenAI’s public release of GPT-5.5 (offering “Mythos-like hacking” capabilities widely). The piece emphasises that AI is moving beyond language games to enterprise process automation through shared team agents.

The structural story is agent runtimes / harnesses-as-a-service productizing — exactly the HaaS framing Osmani names ten days later. Anthropic ships Claude Managed Agents; OpenAI decouples its agent harness and releases it open-source via the Agents SDK; Amazon launches agent registry within AWS Bedrock AgentCore; Cursor 3 repositions as agent orchestrator, demoting the IDE to secondary role. Four vendors converging on the same productization within a single month.

TL;DR — the month’s standout threads

• AI becoming operational. AI is moving beyond language games to enterprise process automation through shared team agents. The framing line of the month.
• Claude Mythos under restricted access; GPT-5.5 public. The capability-gap-and-safety-policy split between Anthropic and OpenAI is now public knowledge. AISI analysis: Mythos “represents a step up over previous frontier models.” But small open-weight models can match Mythos performance at vulnerability discovery — the capability gap may be narrower than the safety-policy gap suggests.
• Workspace agents become productized. OpenAI workspace agents (team-wide tool sharing and automation); Anthropic Claude Managed Agents (prebuilt harness infrastructure for agentic workflows); Anthropic Claude Code routines (scheduled/triggered packages with prompts and connectors); OpenAI decouples harness → Agents SDK open-source; Amazon agent registry in Bedrock AgentCore; Cursor 3 as orchestrator, IDE secondary. Five vendor-side moves toward HaaS in a single month.
• Stanford 2026 AI Index Report (400+ pages) published this month — the wiki’s standalone ingest is 2026-04-30-ai-index-report-2026.
• DeepSeek V4 preview — “large model (over 1T parameters)” with frontier-level performance.
• npm supply-chain attack allowed credential theft and self-propagating malware insertion. Kyber ransomware gang transitioning to post-quantum encryption on Windows systems. Two structural security shifts in one month.
• GPT-Rosalind — OpenAI’s biology-tuned model for 50 common laboratory workflows with skepticism bias. First wiki source on a science-domain-tuned frontier model with bias baked in by design.
• Robotics inflection. Boston Dynamics Spot reads gauges and thermometers using Gemini Robotics-ER 1.6. MLB deploys robotic system for ball/strike challenge rulings. Robotics and embodied AI moving from demo to production deployment.
• Anthropic’s 3.5 GW compute deal with Google and Broadcom — power consumption specified rather than chip counts. Direct convergence with Apollo-comparison thesis about AI-foundation capex scale.

Sectional summary

AI Models. Google publishes a catalogue of “1,302 real-world use cases for generative AI”; OpenAI launches GPT Images 2 (image-reasoning before generation); Anthropic demonstrates Claude supervising alignment research, outperforming human researchers; Moonshot Kimi K2.6 + Vendor Verifier tool; Alibaba Qwen3.6-35B-A3B MoE (3B active params); Stanford 2026 AI Index Report (400+ pages); DeepSeek V4 preview (>1T params); OpenAI GPT-5.5 (reduced token counts; increased hallucination rates); Google Gemma 4 (iPhone/Android on-device).

Software Development. OpenAI workspace agents; Microsoft Critique and Council (combining Claude and GPT for research problems); Stash (open-source memory layer); Anthropic addresses Claude Code performance regression (detailed postmortem); Glif agent (multi-LLM unified access with auto-selection); OpenAI decouples agent harness → Agents SDK open-source; Anthropic Claude Code routines (scheduled/triggered packages); Anthropic Claude Managed Agents (prebuilt harness); Amazon AWS Bedrock AgentCore agent registry; Cursor 3 repositioned as agent orchestrator.

Security. Firefox 150 patches Tor de-anonymization vulnerability; Kyber ransomware → post-quantum encryption on Windows; npm supply-chain credential theft + self-propagating malware; Signal iOS notifications vulnerability (environmental, not protocol); Google Security Operations adds three AI agents (Threat Hunting, Detection Engineering, Third-Party Context); Microsoft reports Teams help-desk impersonation attacks; NIST stopped scoring lower-priority CVEs; Anthropic Project Glasswing restricts Claude Mythos to critical-infrastructure use; AISI Mythos analysis; small open-weight models match Mythos at vulnerability discovery.

Infrastructure and Operations. Microsoft allows enterprise admins to uninstall Copilot; Google 8th-gen TPUs (8t for training, 8i for inference); Google open-sources Scion (agent-orchestration testbed); Anthropic 3.5 GW compute deal with Google + Broadcom (power specified, not chip counts); Ollama adopts Apple MLX for Apple-silicon performance.

Web. “Who Will Maintain the Web When PHP’s Veterans Retire?”; Laravel injecting cloud-service ads into agents; IPv8 proposal (full IPv4 backward compatibility + security + address depletion); Cloudflare EmDash (WordPress alternative); Is BGP Safe Yet? web app (ISP routing protocol implementation test).

Biology. OpenAI GPT-Rosalind (biology-tuned model, 50 common lab workflows, skepticism bias baked in).

Robotics. Boston Dynamics Spot reads gauges + thermometers via Gemini Robotics-ER 1.6; MLB deploys robotic ball/strike challenge ruling.

Cross-positioning with the wiki

• HaaS news-side companion to Osmani’s harness-engineering essay. Osmani’s HaaS section names Claude Agent SDK, Codex SDK, OpenAI Agents SDK as the substrate trio. This digest reports OpenAI decoupling and open-sourcing its harness and Anthropic shipping Managed Agents — the productization of HaaS in real time.
• Empirical-anchor companion to Anthropic Managed Agents source page. The Anthropic primary source ingest from 8 April; the digest names it again as the month’s top runtime story.
• AI-foundation capex scale companion to MGI 2026. Anthropic’s 3.5 GW power deal is the first wiki source where compute-deal capacity is specified in power-consumption terms rather than chip counts — a structural shift toward the power-as-bottleneck framing MGI’s 2040-projection scenario names.

Named entities (this ingest)

• Mike Loukides — see January. Also Claude as co-author on this issue (the wiki does not have an “AI-coauthor” entity convention; treating Claude as a tool-credit rather than an entity-page subject).
• O’Reilly Media — see January.

Plus dangling first-mentions of: GPT-5.5, GPT Images 2, Claude Mythos, Project Glasswing, AISI (AI Security Institute), Kimi K2.6, Qwen3.6-35B-A3B, DeepSeek V4, Gemma 4, Claude Managed Agents (entity-eligible: ingested separately), OpenAI Agents SDK, Cursor 3, Amazon Bedrock AgentCore, Kyber ransomware gang, GPT-Rosalind, Boston Dynamics Spot, Gemini Robotics-ER 1.6, MLB, IPv8, EmDash, Glif, Stash, Plumb (April carry), Critique and Council.

Source-quality notes

Same as January (single secondary-summary source; confidence 0.65), with an additional provenance note: this issue is co-authored by Mike Loukides and Claude. The wiki treats this as transparency from the publisher about AI-assisted curation, not as a downgrade of the digest’s authority.