This week, AI showed three very different faces. OpenAI shipped a cybersecurity model so capable it earned the company's first Critical rating in that category. An independent evaluator found that AI is now speeding up the research that builds the next generation of AI. And Meta turned its assistant into something you can talk to face to face, with a voice and an avatar that reacts while you speak. Here is what happened, with every claim verified against primary sources.
Key Takeaways
OpenAI launched GPT-6 Cyber on November 5, scoring 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and earning a Critical rating under OpenAI's Preparedness Framework. It found 81 previously unknown high-severity vulnerabilities in Firefox and built working exploits for 948 of 1,724 reproduced Linux kernel vulnerabilities. (OpenAI)
METR's independent assessment found Anthropic's AI R&D capability is improving about twice as fast as it did from March 2024 to October 2025, with roughly 80% of the current acceleration attributed to using AI in R&D itself. The assessment ran from August 12 to October 29, 2026. (METR)
Meta rolled out real-time voice and expressive avatars for its Muse assistant in the US, on mobile, web, and Mac, powered by Muse Realtime Voice and Muse Realtime Avatar, with hundreds of voice options and three presence modes. Voice on AI Glasses and Muse Charm is coming soon. (Meta Newsroom)
1. OpenAI's GPT-6 Cyber Earns the First Critical Cybersecurity Rating
OpenAI on November 5 launched GPT-6 Cyber, a model built specifically for cybersecurity work, and the launch came with a milestone the company had not crossed before. Under OpenAI's Preparedness Framework, the model is rated Critical in cybersecurity, up from High for its predecessor, GPT-5.6 Cyber, reports OpenAI.
The headline number is the ExploitGym score: 48.5%, up from 32.1% for GPT-5.6 Cyber. That is about 16 percentage points higher, a 51% relative improvement on OpenAI's own benchmark for offensive security capability. The model also stayed in scope far better than its predecessor, avoiding out-of-scope services in 99.93% of ExploitGym Honeypot tests without deployment prompts, up from 94.49%, and reaching 100% with the prompts. Its web penetration test recall was 40.3%, compared with 32.0% for GPT-6 Astra.
The real-world findings are the part that matters most. GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities in Firefox, as classified by Mozilla, with Mozilla releasing fixes for many of them, and it showed that 30 known issues were more severe than previously reported. On the Linux kernel, it reproduced 1,724 vulnerabilities and developed working privilege-escalation exploits for 948 of them, more than five times the 166 exploited by GPT-5.6 Sol with the same harness. Of those findings, 304 critical ones were disclosed to maintainers with candidate patches.
Access is deliberately limited. GPT-6 Cyber is initially available through the API and Codex only to a select group of approved Daybreak Red customers, expanding to all Daybreak Red customers soon. OpenAI also announced a $1 billion Daybreak for Frontline Defenders grant program to subsidize access for resource-constrained defensive teams. The message is clear: a model this capable at finding vulnerabilities is powerful in the hands of defenders, and risky in the hands of anyone else.
2. METR: AI Is Now Speeding Up AI Research Itself, by About 2x
The second story is not a product launch. It is a measurement of something the AI industry has talked about for years: AI systems helping to build the next generation of AI systems. METR, an independent AI evaluation nonprofit, published an assessment on November 5 of how fast Anthropic's AI research and development is accelerating, and the headline estimate is about 2x, reports METR.
The assessment ran from August 12 to October 29, 2026, about 11 weeks with roughly 2.5 full-time-equivalent staff, and METR had high access to internal data: pretraining and reinforcement learning metrics, training compute figures, internal benchmark scores including unreleased models, AI speedup analyses, forecasts, and staff interviews.
The core finding comes from AECI scores, Anthropic's version of the Epoch Capabilities Index. Since October 2025, those scores have grown at an estimated 1.9x the pre-October 2025 rate, with a 95% confidence interval of 1.5x to 2.8x. The trend steepened from about 15 points per year before October 2025 to about 29 points per year after, which METR describes as roughly two years of historical progress compressed into one year. Subjectively, METR puts it this way: as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate of March 2024 to October 2025, with high uncertainty. METR says it would not be surprised by 1.3x or 3x.
METR's best guess is that a majority, perhaps around 80%, of the current acceleration comes from Anthropic's extensive use of AI to do R&D, though with substantial uncertainty. The acceleration in late 2025 was likely influenced more by non-AI factors. METR is also confident the speedup was not caused by faster scaling of training compute, and it found no clear bottleneck in algorithmic progress, compute, or data that would prevent further progress in the next 12 months, though it is not confident in that bottleneck assessment.
The caveat matters as much as the headline. METR states plainly that Anthropic redacted a majority of the evidence METR found most informative, so the public has to rely heavily on METR's judgment. Staff self-reports of speedup in a Slack poll reached 12x in October 2026, which METR calls very loose evidence only. Even with those limits, this is the first independent, data-backed estimate of AI accelerating AI research, and if the 2x figure holds, the pace of model improvement is itself compounding.
3. Meta Gives Muse a Voice and a Face
Meta on November 5 rolled out real-time voice conversations and expressive avatars for Muse, its AI assistant, starting in the US on mobile, web, and Mac, reports Meta Newsroom. Users start by tapping the phone icon in the top-right corner of the Muse app.
The feature is powered by two new technologies: Muse Realtime Voice and Muse Realtime Avatar. The avatar uses subtle expressions and gestures that match the tone of the conversation and reacts while the user talks. There are hundreds of voices to choose from, from what Meta describes as the energy of a New Yorker to the flair of a pirate, changeable at any time, and users can interrupt Muse mid-conversation. Three presence modes are available: voice only, avatar alongside the conversation, or full screen face to face.
Voice conversations are coming soon to AI Glasses and Muse Charm, with no date given. The feature was previewed at Connect 2026. Early hands-on coverage noted quick replies with occasional pauses while Muse looks things up, and that Muse hears the user but does not see them.
This is a different kind of AI news from the other two stories. GPT-6 Cyber is about capability, and the METR assessment is about speed. Muse's voice and avatar are about the interface: AI assistants are moving from text boxes toward something closer to a conversation with a person. That shift has real consequences for how people trust, use, and relate to these systems, and Meta is betting that a face and a voice will make Muse feel less like a tool and more like a companion.
How the Week's Stories Stack Up
Aspect | Story | Key Number | Source |
|---|---|---|---|
Cybersecurity | OpenAI launches GPT-6 Cyber | 48.5% on ExploitGym, Critical rating, 81 new Firefox vulnerabilities | OpenAI |
AI research | METR assesses Anthropic's R&D acceleration | About 2x faster, 1.9x AECI growth rate, around 80% attributed to AI use in R&D | METR |
AI assistants | Meta adds real-time voice and avatars to Muse | Hundreds of voices, three modes, US rollout on mobile, web, and Mac | Meta Newsroom |
The thread across all three: AI is getting more capable at the hardest tasks, faster at improving itself, and more human in how it talks to us. Last week was about where AI physically lives, in orbit, in China, and on the power grid. This week is about what AI can do, how fast it can get better, and how it shows up when you talk to it.
If you like quick video takes on how new open AI projects work, browse the repo signals for 30-second overviews. For last week's roundup, see This Week in AI: November 4, 2026.
Read this article in Bahasa Indonesia: Berita AI: 08 November 2026.