Radar Trends to Watch: March 2026

Loukides — Radar Trends to Watch: March 2026

The third installment of the monthly digest, published 3 March 2026, covering February 2026. Opening framing: OpenClaw’s explosive growth in February — agent-only social networks and games emerge, multiple security-focused clones appear, and Loukides positions OpenClaw as “the next layer on top of AI agents” per Andrej Karpathy’s assessment.

The month’s structural story is security adjacent to agent infrastructure: the OpenClaw ecosystem produces immediate exploit-bait (HackMyClaw, IronClaw, credential-theft active operations), and the same digest cycle reports AI finding 12 vulnerabilities in OpenSSL via Claude and 500 zero-days discovered via Claude Opus 4.6. The harness-as-attack-surface theme that Osmani’s harness-engineering essay later names is implicit throughout.

TL;DR — the month’s standout threads

• OpenClaw becomes the “next layer on top of AI agents” per Karpathy — agent-only social networks, sandbox-permission challenges, alternatives emerging (IronClaw with stronger sandboxing).
• AI as offensive-security tool, at scale. 500 zero-days discovered via Claude Opus 4.6 analysis; 12 OpenSSL vulnerabilities found via Claude; HackMyClaw gamified prompt-injection challenge launches.
• Agent-to-Human (A2H) protocol surfaces as a category — formalising agent-human handoff as a discrete primitive. Adjacent to the wiki’s agent-development-lifecycle coverage.
• Claude Remote Control — cross-device desktop session continuation. First wiki mention of session-portability as a discrete agent capability.
• C compiler via agents — 100,000 lines of Rust, ~$20K token cost. The most concrete agent writes a serious systems program result in the digest series so far.
• Ransomware evolves from encryption to extortion — the threat-model shift Loukides flags.
• Lyria 3 — 30-second musical clip generation from verbal prompts. Multimodal-AI continues moving up the creative stack.

Sectional summary

AI (10 items). Moonshine Note Taker (local voice transcription); Nano Banana 2 / Gemini 3.1 Flash; Claude Remote Control; OpenClaw sandbox limitations; Alibaba Qwen 3.5; Lyria 3; Agent-2-Human protocol; Claude Sonnet 4.6 + Gemini 3.1 Pro releases; multimodal RAG (enterprise graphs/images); Waymo’s World Model for traffic-behaviour simulation.

Programming (7 items). Anthropic Claude Max — six months free for open-source maintainers; Pi coding agent (minimal terminal-based); C compiler via agents (~$20K in tokens); GitHub PR controls (disable/limit); Showboat and Rodney (agents demo working software); Gas Town (codeless programming exploration continues); macOS sandboxing tools (deprecated → easier alternatives).

Security (14 items, dominant section). Smart-glasses detection app; Agentsh (execution-layer security via bash replacement); attacks on time synchronization (factory infrastructure threats); AI discovers vulnerabilities — 12 in OpenSSL via Claude; HackMyClaw gamified prompt-injection challenge; OpenClaw credential theft active operations; Deno sandbox potential; IronClaw OpenClaw alternative emphasizing sandboxing; malicious coding challenges (hidden npm/PyPI dependencies); 500 zero-days via Claude Opus 4.6 analysis; VS Code malware extensions exfiltrating to China; Bizarre Bazaar (LLM API attacks for resource theft); ransomware evolution (encryption → extortion); personal data removal tools.

Web (3 items). Markdown for Agents (Cloudflare HTML→markdown service); WebMCP (proposed API standard for agent-accessible web apps); Firefox 148 (AI feature opt-out).

Operations (2 items). Babyshark (simplified Wireshark TUI); laser-etched glass (Microsoft long-term data storage).

Things (2 items). Desk robots; Doom on Lego.

Cross-positioning with the wiki

• AI-as-offensive-security complements responsible-ai coverage with a new dimension — the same capability that finds 500 zero-days legitimately can also be turned malicious; this is the dual-use thread that will recur in April and May digests.
• OpenClaw’s emergence as a category is the wiki’s strongest signal that persistent always-on agent memory is becoming infrastructure. Worth tracking for concept-page promotion once a second non-Loukides source treats it.
• Anthropic Claude Max for OSS maintainers is an OSS-economics datapoint that complements Masad’s “only two jobs left” AI-native-company framing.

Named entities (this ingest)

• Mike Loukides — see January.
• O’Reilly Media — see January.

Plus dangling first-mentions of: OpenClaw, IronClaw, HackMyClaw, Agentsh, Claude Sonnet 4.6, Gemini 3.1 Pro/Flash, Lyria 3, Waymo World Model, WebMCP, Agent-2-Human protocol, Pi coding agent, Showboat, Rodney, Babyshark, Moonshine Note Taker.

Source-quality notes

Same as January: single secondary-summary source; confidence 0.65.