This roundup covers October 17–24, 2026 — the week in AI as we published it, grouped by topic, with every story linked to our full coverage and the original sources. One entry per development; no padding.

Agents & products

Claude Duet launched as a proactive agent app (October 21). Anthropic's new app makes Claude initiate and complete work rather than wait in a chat — available as a research preview for Pro and Max. Why it matters: the first consumer-facing proactive agent from a major lab, gated to power users while governance questions get answered. Full coverage →

ChatGPT locked chats began rolling out for Business (October 20). Conversations protected separately from regular history, across web, iOS, Android, and desktop. Why it matters: the first practical privacy control for the most common workplace AI mistake — pasting sensitive material into chats. Full coverage →

Gemini Agent continued its business rollout (from October 8). Google's universal work agent operates across Gmail and connected apps for Google Cloud customers. Why it matters: the biggest test yet of agent governance inside real enterprise tooling. Full coverage →

Models & platforms

The Decisions API with GPT-6 Luna went generally available (October 22), with expanded per-request capabilities — production support for structured, justified AI decisions. Why it matters: GA turns an interesting preview into something you can build a business on. Full coverage →

GPT-5.5 retired from consumer surfaces (October 14) — removed from ChatGPT, ChatGPT Work, and Codex on all plans; the API unaffected. GPT-6.1 Sol now handles complex work. Why it matters: default outputs changed for anyone who tuned prompts around GPT-5.5. Full coverage →

Gemini 4 Argon became the app default, with a native Windows app (October 13) — closing Gemini's last major platform gap. Full coverage →

Policy & business

Meta delayed its AI training terms (October 23). The commercial terms — set to allow AI training on business-linked pages from October 30 — were paused after publisher and advertiser backlash, per Reuters. Why it matters: the data fight between AI labs and publishers is now a negotiated, public process. Full coverage →

Anthropic published its 2026 Usage Policy update (October 8, effective November 12), paired with a public consultation on Claude's values. Why it matters: usage policies are enforceable documents; developers have three weeks to review. Full coverage →

AI funding kept setting records: $102 billion in Q3 — 64% of global venture funding per Crunchbase — with October tracking at $25.7 billion across 59 rounds. Why it matters: two-thirds of the world's venture dollars now go to AI, and the application layer is getting financed at sane sizes alongside the giants. Full coverage →

Security

Indirect prompt injection is now an operational attack, per a Cloud Security Alliance research note — with Google reporting a 32% relative rise in malicious injection content across billions of crawled pages. Why it matters: the shift from lab demos to real exploitation changes what every team deploying agents must do this quarter. Full coverage →

Coverage window: October 17–24, 2026. This roundup lists only developments we published full coverage of this week — earlier October stories (Gemini Agent, Haiku 5.5, Copilot's multi-model update) are in the linked articles.