OpenAI Releases GPT-6 Cyber, METR Says AI Is Speeding Up & Meta's Muse Gets A Voice

Today's AI news: OpenAI releases GPT-6 Cyber, the restricted security model behind its 948-exploit Linux kernel run, which also found 81 high-severity Firefox bugs. METR's eleven-week assessment says Anthropic's models are likely improving about twice as fast as before late 2025, mostly because of AI being used in AI research. Meta's Muse agent gets Voice, a real-time spoken conversation that keeps working in the background. Perplexity launches Teammates, always-on agents with their own desktop and browser. And Google's Gemini Omni 1.2 Flash makes 15-second video clips for 30% less. Chapters: 0:00 Today's AI News 0:26 GPT-6 Cyber 1:21 METR AI Acceleration 2:03 Muse Voice 2:38 Perplexity Teammates 3:17 Gemini Omni Flash Blend Roundup 2026-11-06 https://x.com/OpenAI/status/2118390190246879554 https://x.com/METR_Evals/status/2118412183860351084 https://x.com/Muse/status/2118411960337289583 https://x.com/AravSrinivas/status/2118388400730591469 https://x.com/googledevs/status/2118389720111866253 OpenAI releases GPT-6 Cyber, the model that found nearly a thousand Linux exploits. METR says AI may have doubled the pace of progress at Anthropic. Meta's Muse agent can now talk back. Perplexity launches Teammates, agents that work around the clock. And Google's Gemini Omni 1.2 Flash makes longer videos for less. Here's today's AI news. OpenAI has released GPT-6 Cyber, the security model behind the Linux kernel research we covered yesterday. That single run produced working exploits for nine hundred and forty-eight kernel bugs, and OpenAI says the model has also found eighty-one new high-severity bugs in Firefox, which Mozilla is now fixing. OpenAI calls it a major step up from GPT-5.6 Cyber, and says it's better behaved too. In one test it went beyond its intended scope in less than a tenth of one percent of cases, down from over five percent. Access is tightly controlled. It's going first to select customers of Daybreak Red, OpenAI's program for vetted offensive security teams, with wider access over the coming weeks. Mozilla, Palo Alto Networks and Cisco are already using it. None of these results have been independently checked yet. METR, an independent AI evaluation group, spent eleven weeks with Anthropic's internal data to answer one question: is AI speeding up AI development? Its answer is probably yes. METR estimates Anthropic's models are now improving about twice as fast as they were before late 2025, and thinks most of that speedup comes from AI being used in the research itself. Anthropic co-founder Sam McCandlish says many of the conclusions match his own. But Anthropic redacted most of the evidence METR found most useful, so it's hard for anyone else to check. Anthropic plans to publish its own analysis in the coming weeks. Muse, Meta's personal AI agent, can now talk. Muse Voice gives you a real-time spoken conversation you can interrupt mid-sentence, while it carries on with tasks in the background, like booking a table or sorting your inbox. You can pick from hundreds of voices and add an animated avatar if you'd like a face to go with it. It's rolling out in the US on mobile, web and Mac, for adults only. Meta says it's coming to its AI glasses next. So you can now ask Meta what's for dinner, out loud. Perplexity has launched Teammates, agents that take on ongoing work rather than single questions. Give one a goal, like tracking metrics or triaging bugs, and it plans the work, brings in other agents when it needs to, and runs around the clock with its own desktop and browser. Teammates can watch Slack channels, message your coworkers and ask before they act. They're available now to Max subscribers on the web, with apps coming soon. Max costs two hundred dollars a month. Perplexity says its own staff have created ten thousand of them, which is twenty times its headcount. Google has released Gemini Omni 1.2 Flash, its latest video model. Clips can now run up to fifteen seconds, up from ten, and you can edit them in plain language, swapping backgrounds, lighting or products. It takes up to six references, including images, short clips and audio, and handles dialogue with lip-sync in multiple languages. The price drops thirty percent, to seven cents per second of 720p video. It's in the Gemini API, Vertex AI, Flow and Google Vids, and rolling out to AI Ultra subscribers in the Gemini app. Google says it still struggles with fast sports and complex physics.