Latest AI news — September 23, 2026

Google: Gemini 3.8 Flash TTS and Flash-Lite TTS. Google launched two new text-to-speech models: Flash TTS designs new character voices from natural-language prompts, with line-by-line performance control, in 100+ languages; Flash-Lite is tuned for high-volume dubbing and voice agents. More than 2,000 ready-made voices, voice replication from a 30-second sample (with consent checks), and SynthID watermarking on every generated clip. Available in Google AI Studio, the Gemini API, Gemini Notebook, and Google Vids; the Flash-Lite enterprise API is coming soon. Source: Google.

UN Security Council: AI leaders brief in New York. Under France’s presidency, the Council is debating AI and international security; briefers include Sam Altman (OpenAI), Dario Amodei (Anthropic), Clément Delangue (Hugging Face), and Yoshua Bengio (UN scientific panel). OpenAI says Altman is urging shared capability and safeguard benchmarks; ahead of a Trump–Xi meeting, Washington and Beijing are also discussing an AI-incident notification system. Source: Reuters, Globe and Mail / Reuters, UN Web TV.

Cisco Talos: CLOSEDQUORUM — first documented autonomous AI C2 implant. Talos tracks AI-integrated malware with its open-source CAIRN toolkit; the first case study is a Windows Go implant that, after install, queries up to four commercial LLMs (DeepSeek, Qwen, Mistral, Gemini), picks the next step by majority vote (steal / inject / persist), and needs no continuous human C2 server. Targets: LSASS, browser passwords, crypto wallets; exfil over a Discord webhook. The public build uses dummy keys — no in-the-wild campaign confirmed. Source: Cisco Talos.

Meta: Muse is “strongly OpenClaw-inspired.” Nat Friedman (Meta Superintelligence Labs) told TechCrunch and confirmed on X that Muse was built from scratch but is “definitely heavily inspired” by open-source OpenClaw — the team, he says, “nailed” Peter Steinberger’s harness, including similar workspace filenames and near-identical SOUL.md content. The goal: an OpenClaw-like agent at billion-user scale, safer and simpler. Source: TechCrunch.

OpenAI: internal model on 100+ open math problems — plus an independent advisory group. An internal model launched August 28 solved the Navier–Stokes Millennium problem and more than 100 long-open math questions across most areas. The company is coordinating evaluation and communication with an independent Advisory Group on Mathematics and Artificial Intelligence forming at the Institute for Advanced Study (including Gowers, Hairer, Witten) — the group does not advise internal R&D pace and works without OpenAI pay. Source: OpenAI.