Anthropic’s AI is helping build better AI faster. The researchers checking its progress say the public can’t see much of their evidence.

Welcome, humans!

It’s been a wild 24 hours for AI, from models that can hunt for security vulnerabilities to chatbots that can examine your bank transactions. And while tech companies keep introducing new AI capabilities, researchers are working out how quickly those capabilities are improving. Today’s news offers a closer look at both sides of that race.

🆕 NEW From The Neuron

• OpenAI found 948 ways to exploit Linux. Our latest analysis explains how AI is helping researchers find vulnerabilities and giving software maintainers a chance to fix them.
• Claude investigated a 40-year-old cell mystery. Researchers used Claude to shorten a three-day lab process to 90 minutes, although the model’s mistakes also delayed the investigation.
• OpenAI changed its data center pitch. The company is promising greater transparency with local officials, including an end to mandatory nondisclosure agreements.

Around the Horn: Friday, November 6, 2026

Anthropic’s AI development is accelerating, but much of the evidence is hidden

An independent assessment found Anthropic’s AI capabilities are improving at roughly twice the rate seen during an earlier period, and researchers believe AI itself is responsible for much of the acceleration.

In a November 5 report, independent research organization METR estimated that Anthropic is now making progress on AI capabilities at approximately twice its previous pace. Researchers examined internal development data from August 12 through October 29. They found evidence that AI tools are helping Anthropic employees work faster on tasks such as coding and analyzing experiments.

Here’s why the finding is interesting: If AI helps researchers build better models, those models could eventually develop their successors even faster. Researchers call this recursive self-improvement. Most of the acceleration METR observed still came from AI assisting people rather than independently conducting entire research projects, but it found no clear bottleneck that would necessarily prevent further acceleration over the next year.

The public, however, cannot examine much of the evidence behind those conclusions. METR said Anthropic redacted most of the material it considered most informative in its roughly 100-page assessment. Anthropic said the redactions were intended to protect confidential information that could help competing AI developers.

METR could publish its overall conclusions, but many underlying measurements and forecasts remained confidential. So the researchers are warning that AI progress may be speeding up, while asking the public to trust conclusions it cannot fully verify.

The next question is whether increasingly capable models will begin completing larger portions of AI research independently, and whether outside evaluators will have enough access to assess the consequences.

🏆 TOP 5 NEWS (Around the Horn)

• AI stocks plunged as investors questioned whether the infrastructure boom could continue at its expected pace. Political resistance intensified after the U.S. midterm elections, while governments from New York to Amsterdam tightened restrictions on data center development. Meanwhile, Google announced a $17 billion investment in a Wyoming data center, which will rely partly on a planned 702-megawatt fuel-cell development.
• The Justice Department launched a Super Intelligence Task Force to investigate unlawful uses of advanced AI. The new unit will focus on threats from foreign governments and criminal organizations, alongside fraud involving AI technologies. The department says it also wants companies to maintain safeguards that help prevent misuse.
• Airbnb plans to launch an AI travel agent in the first half of 2027. CEO Brian Chesky says the experience will assemble maps, photos, and travel comparisons around a user’s request instead of relying exclusively on text chat. The agent will also help guests change reservations or resolve booking problems during their stays.
• USA Today blamed AI search for part of its declining traffic and digital revenue after reporting disappointing third-quarter results. CEO Mike Reed pointed to changes in how people search for information, while the publisher is also exploring AI-supported advertising and video opportunities. Publishers face pressure to replace traffic and revenue as search behavior changes.
• Military officials reported mixed results from AI deployments. At a November 5 defense technology event, U.S. Army and Air Force officials described gains in tasks such as automated target recognition. However, broader intelligence analysis has delivered inconsistent results, and some complex decisions still require substantial human work.

Honorable Mentions

• OpenAI collaborates with Australian universities to bring ChatGPT into teaching and research. In an OpenAI-commissioned survey of more than 1,000 Australian students, 81% reported using AI several times a week, while 28% said they had already built an AI agent or assistant.
• Researchers are using AI to study dolphin communication. Wild Dolphin Project founder Denise Herzing described how scientists are analyzing decades of recordings to identify patterns in dolphin sounds. The team has also experimented with underwater devices that play specific sounds, although researchers have yet to demonstrate a genuine two-way conversation.

🍪 TOP TREATS TO TRY

• OpenAI introduced GPT-6 Cyber for advanced cybersecurity work. The specialized model lets authorized researchers identify vulnerabilities and develop exploits to test security defenses. OpenAI classified its cybersecurity capabilities as Critical and restricted initial access to selected Daybreak Red organizations. MTS also reported the launch after OpenAI disclosed AI-powered Linux security research.
• Claude now connects to personal bank accounts through its Personal Finance beta. U.S. Pro, Max, and Ultra subscribers can connect selected accounts through Plaid and ask Claude about spending, upcoming bills, or unusual transactions. Access is read-only, so Claude cannot transfer or invest money. The service can review up to two years of transaction history, and excludes conversations within the Personal Finance hub from model training.
• Meta released real-time expressive avatars for its Muse assistant in the U.S. Users can speak with Muse via mobile, web, or Mac app while an animated avatar responds with facial expressions and gestures. The assistant supports interruptions and offers hundreds of voice choices, with support for AI glasses and Muse Charm planned later.
• Perplexity launched Teammates, its always-on work agents. The agents can handle multiple assignments, monitor changes, and collaborate on tasks. They can also use existing Perplexity Computer projects and permissions, and users can inspect or redirect their work. Teammates initially launched for Max subscribers on the web.
• Mistral expanded Studio with managed AI agents and enterprise apps. Developers can connect agents to company systems, monitor their actions, and automate workflows. Apps and Managed Agents are in public preview, while Workflows is available on Mistral Cloud. Private cloud and on-premises deployment are planned for late November.
• Cursor now builds and deploys live web apps from natural-language instructions. Its cloud agent can create an app, provision a database, and configure authentication before deploying it to a shareable link. Security checks run before publication, and users must explicitly approve deployment. The capability is available on paid plans with cloud agents enabled.
• Google released Gemini Omni 1.2 Flash for AI video generation and editing. The model supports clips up to 15 seconds, up from 10 seconds in the previous version, and allows iterative edits through follow-up prompts. Google reduced pricing by 30% to $0.07 per second for 720p video, and improved consistency across characters and reference images.
• Planet introduced an AI app that lets users explore satellite imagery using natural language. Users can search approximately a decade of Earth observations to investigate changes in locations over time. The map-based tool is available, with free access offered through Planet’s website.
• Thinking Machines Lab released Inkling 2 Small, a model designed for coding, visual tasks, and autonomous work. It has 594 billion total parameters, with about 17 billion active per request. Developers can access it through Tinker, while the company plans to release downloadable model weights after additional safety testing.
• Baidu released ERNIE 5.5 Flash Base, an openly available 280-billion-parameter AI model. Only 14 billion parameters are active per input, reducing computation required per request. Baidu deliberately released the base model before instruction fine-tuning so that developers can customize it for their own apps.
• Glasshand gives AI agents access to real iPhones and Macs hosted in its data centers. Developers can rent physical devices for agents such as Claude Code or Codex to test apps and perform device-based tasks. Glasshand says its resets the devices after each rental.
• Rerun lets users create always-on AI agents without coding. Agents can monitor inboxes, follow up on invoices, or complete recurring business tasks. Users can inspect what they do and require approval before sensitive actions. The latest launch also introduced an open-source, self-hostable option.

Previous Around the Horn Digests

Catch up on the stories you missed:

• Thursday, November 5, 2026: OpenAI’s Linux vulnerability discoveries, Anthropic’s cell research and Apple’s AI spending cuts.
• Wednesday, November 4, 2026: OpenAI’s security scrutiny, Google’s classroom AI dispute and concerns about autonomous agents.
• Tuesday, November 3, 2026: Anthropic’s self-improvement warning, deceptive deepfake apps and election misinformation preparations.
• Monday, November 2, 2026: OpenAI’s upcoming subscription changes, Meta’s advertising growth and AWS’s latest results.
• Friday, October 30, 2026: OpenAI’s safety-monitoring experiments, Reddit’s AI search concerns and Canada’s support for Cohere.
• Thursday, October 29, 2026: Gemini 4 Carbon’s cybersecurity capabilities, a lawsuit involving AI medical advice and GitHub’s AI coding data.
• Wednesday, October 28, 2026: Trump’s remarks about AI terminology, OpenAI’s reported Google TPU plans and Anthropic’s government expansion.

That’s a Wrap

That’s 21 developments drawn from 24 sources, covering everything from Anthropic’s accelerating research to AI-powered dolphin studies. If you made it this far, congratulations. You survived an entire AI news cycle, which may soon require its own AI assistant.

For your daily AI update in five minutes, subscribe to The Neuron.

See you next time!

P.S. Know someone who spends their morning catching up on AI announcements? Forward this to them and tell them to subscribe here.