OpenAI introduced GPT-6.2 Astra and Sol on 17 November 2026. It claims better professional outputs, coding, computer use and efficiency at lower cost across ChatGPT Work, Codex and the API.
Cursor says the trojanized Anthropic Claude Code extension on its extension marketplace came from a stolen access key. The key had been issued to Fireworks AI, exposed in a Fireworks breach, and used to copy data from a read-only store.
Google launched Gemini 4 Neon on November 18, a mid-tier model between Hydrogen and Carbon aimed at coding, agentic and analysis work, with a 1M-token context and four thinking levels. API pricing is $3 per million input and $15 per million output tokens; it is in the Gemini app for AI Pro/Ultra and in Google Antigravity. Neon replaces Argon in the picker and Gemini 3.1 Pro is being retired.
Anthropic says Claude Fable 5.2 solved two open problems in black hole physics, including completing a 2003 Hawking-Page phase-transition calculation and deriving an exact modular-bootstrap bound on the lightest black hole mass in 3D quantum gravity. The bootstrap results were formally checked in Lean, and a physicist reviewed every main result.
Anthropic launched Claude Meetings in beta on November 12: Claude can join Google Meet calls as a visible participant, answer questions, build docs and whiteboards live, and write notes, action items and recaps. It is rolling out to Team and Enterprise plans; organisations with zero data retention, HIPAA-ready setups or customer-managed keys are excluded. Transcripts and summaries, not audio, are saved per meeting in claude.ai.
Google introduced Gemini 4 Neon, adding to the Gemini 4 line-up of Carbon, Argon and Hydrogen. Neon costs $3 per million input and $15 per million output tokens, and the Carbon and Argon discount prices double on 1 December.
Google reports threat actors now use autonomous multi-agent frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to run multi-stage intrusions, including building exploits from software patches.
OpenAI's GPT-Live 1.1 full-duplex voice model is now on Vercel AI Gateway; it can be interrupted mid-reply and transcribes more accurately than the previous version. Existing GPT-Live integrations only need the new model ID, and client delegation lets it hand tasks to any text model on the gateway.
Claire Vo, who runs the AI product ChatPRD, reports organised token fraud: groups cycle free trials and promo codes to resell tokens or run commercial workloads through her agent. Simon Willison notes the abusers also probe which underlying model is in use. ChatPRD is ending free trials entirely.
Researchers from Waterloo, Zhejiang University and KTH found that AI agents retrained on larger sets of prompt-injection attacks became easier to hijack in three of four tests. The automated red-teaming tool that found more working attacks produced the weaker defence, because it kept finding near-duplicates of the same injection.
GitHub published a practical guide to agent skills: reviewable files that capture repeatable engineering work and run on any agent host, while tools, project context and permissions stay host-specific.
Simon Willison and Prime Radiant describe an external sandbox pattern, where code from coding agents runs in a separate sandbox while the agent stays outside. Three server implementations are released, with an HTTP allow-list proxy and credential replacement, usable with Claude Code, Codex and Pi.