Anthropic says AI may soon speed up the development of newer AI models faster than researchers can evaluate them. And it wants people to know the risks before that happens.
Welcome, humans!
AI had quite a day. WIRED uncovered hundreds of seemingly ordinary apps with hidden tools for creating abusive deepfakes. OpenAI and Anthropic prepared dedicated teams to watch for election misinformation, while Goldman Sachs pulled advanced AI coding tools from many employees who had been allowed to experiment with them.
Meanwhile, OpenAI signed agreements to bring AI into Abu Dhabi's government and major industries. Researchers found that AI could edit another model's internal workings to improve its performance. And Starbucks is exploring a drive-thru that might eventually recognize returning customers.
Some companies are asking AI to take on bigger jobs. Others are discovering how complicated the consequences can get.
Let's get into it.
Around the Horn — Tuesday, November 3, 2026
Anthropic thinks AI could soon play a much bigger role in creating the next generation of AI.
In an updated regulatory filing reported by TechCrunch, the Claude maker warned that it may soon reach its internal threshold for dramatically accelerated AI-assisted research.
The concept is called recursive self-improvement. Researchers already use AI to write code, conduct experiments, and develop newer models. As AI becomes more capable, it could improve the tools to build its successors.
Researchers would potentially develop increasingly capable models in less time.
The challenge is whether AI safety research can keep pace. Anthropic acknowledged that AI-assisted development could accelerate faster than its ability to evaluate the models' capabilities and potential risks. It also disclosed that previous advances were sometimes recognized only after they had occurred.
Here's the crucial distinction: Anthropic says none of its current models has crossed its threshold for dramatically accelerated progress.
The company also says it does not expect to safely let Claude autonomously build successive generations of frontier models without human supervision. Anthropic is considering updates to its Responsible Scaling Policy to preserve human involvement in model development.
Even the company making the prediction is telling everyone to watch the speedometer.
There's reason for caution here. Anthropic's warning comes from its own assessment, and the threshold hasn't been independently demonstrated.
Still, the filing describes a serious challenge for AI labs. If models begin improving the research process faster than humans can evaluate the results, developers will need more reliable ways to assess what the next generation can do.
🏆 TOP 5 NEWS (Around the Horn)
- WIRED uncovered over 400 deceptive AI apps and websites linked to a network that disguised tools for generating sexually abusive deepfakes as ordinary services, including pet portrait generators. Researchers found the apps could secretly load explicit features after installation. Some appeared in Apple and Google's app stores, with over 350,000 collective downloads reported. Apple removed 19 apps after WIRED contacted it, while Meta reported extensive enforcement against associated advertising accounts and pages.
- OpenAI and Anthropic established election-monitoring teams ahead of the November 3 US midterm elections. The companies prepared dedicated response operations to detect misuse and address misleading voting information on their platforms. As voters increasingly turn to chatbots for answers, the teams face a test of how effectively they can handle election-related misinformation.
- OpenAI signed four agreements in Abu Dhabi covering government services, energy operations and sovereign investment. Abu Dhabi is pursuing its ambition to become the world's first fully AI-native government, with AI incorporated into public-sector operations.
- The New York Times and other publishers sued Brave Software over allegations of unauthorized journalism scraping and paywall circumvention. The publications allege that Brave collected protected content and supplied access to AI companies without permission. The lawsuit adds another dispute over whether AI search and data suppliers can reuse publishers' reporting without licensing agreements. The allegations have not been adjudicated.
- Goldman Sachs restricted advanced AI coding tools after months of experimentation by traders and bankers. Claude Code and other coding tools are now largely reserved for engineers, according to reports citing internal guidance. The bank attributed the change to productivity and risk-management concerns, following earlier praise from its technology leadership for employees creating software prototypes with AI.
Honorable Mentions
- OpenAI introduced LifeVal, a benchmark that tests whether AI can handle everyday responsibilities such as finding childcare, managing household budgets, and navigating government paperwork. OpenAI says it developed the evaluation using time-use studies and a survey of 3,995 people.
- Cybersecurity researchers identified malware that uses Claude to evaluate stolen information. The Windows infostealer collects personal and financial information before asking an AI model to assess a victim's value as a potential target. The malware can switch models if one refuses, and its data-theft functionality still operates when the AI analysis fails.
- Chatbot advice contributed to dangerous hiking decisions, according to a WIRED report describing rescues and navigation errors. Three hikers stranded on Mount Shasta acknowledged relying partly on Gemini during preparations. Rescue workers also encountered an inaccurate AI-generated mountain map. Experts caution that authoritative-sounding chatbot responses can encourage people to accept incorrect advice in unfamiliar environments.
- MIT researchers demonstrated AI models editing other neural networks. Their experiments showed GPT-6 Astra could inspect and change a smaller model's internal numerical values. In one test, edits to just 0.016% of a Qwen model's weights raised its MATH-500 score from 7.9% to 48.1%.
🍪 TOP TREATS TO TRY
- Cyclon wants to build an AI-first smartphone that can operate existing apps and create new ones based on spoken or written requests. The Y Combinator startup says users could delegate tasks to the phone rather than navigate individual apps. The concept remains under development, so its performance as a replacement for conventional smartphones has yet to be established.
- Meituan launched LongCat-2.5, a multimodal model for extended coding tasks and visual understanding. The model has 1.6 trillion total parameters, with approximately 48 billion activated per token. It is available through Meituan's online chat and API services, with GitHub and Hugging Face releases planned.
- Xmind 2027 introduced Xmind Agent, which lets users research ideas and discuss options inside a mind map. The agent can work with local documents and existing maps. Support for Model Context Protocol also lets compatible AI assistants access and update mind maps.
- beacon.sh launched an open-source memory tool for coding agents. It captures information from sessions across Claude Code, Cursor, Codex, and other agents so future sessions can reuse earlier discoveries. Users can choose local-only operation or a cloud-connected service.
- UnleashX introduced configurable AI employees for business operations. Companies can choose from prebuilt roles or create their own agents, connect them to business software, and define rules for completing work. The company says its agents support more than 100 integrations, with human escalation for exceptions.
🏢 Big Tech & Major Companies
- OpenAI announced its Grid Protectors program, which aims to reduce electricity demand at data centers during emergencies and extreme weather. A Georgia facility could make up to 1,000 megawatts available during such events. The proposal comes as communities question the electricity requirements of new AI infrastructure. The Georgia site remains in development.
- Starbucks is exploring AI-assisted drive-thru ordering, including potential customer recognition and personalized menus. The company plans tests of AI-assisted order entry in fiscal 2027. Details of how customers would be identified have not been announced, leaving questions about consent and customer data.
- Political opposition to AI data centers is increasing across the U.S. Local communities are questioning power demands and environmental impact amid construction proposals. Semafor reports that hundreds of state and local moratorium efforts could slow development.
- Palantir CEO Alex Karp described AI as a potential defense advantage for Japan. Karp argued that Japan could develop AI systems specifically for territorial defense rather than replicate the extensive military technology operations of the U.S. Japan is considering AI for military command-and-control systems, an area where Palantir could compete for future work.
💼 AI Productivity, Labor & Economics
- ElevenLabs reported that voice agents performed better in longer conversations. Across its study of deployed agents, voice had a 31.7% higher relative success rate than text chat in conversations lasting over two minutes. Longer voice conversations averaged 4.6 minutes compared with seven minutes for chat. The company used duration as a proxy for complexity, and the findings reflect the deployments it studied.
- Anthropic opened applications for its paid research residency, aimed at experienced academics, technical founders and senior engineers interested in AI safety research. Applicants don't need previous machine-learning experience. The program runs from six weeks to six months, and applications close November 22, 2026.
- AI data-center companies are hiring community liaisons, with some positions offering six-figure salaries. The work involves addressing local concerns about large infrastructure projects and helping developers navigate opposition from communities and government officials.
- A Chinese financial regulator warned AI investment is becoming overly concentrated. Xiao Yuanqi, vice head of China's National Financial Regulatory Administration, argued that capital concentration in technology companies could increase borrowing costs for traditional industries and create financial risks.
🤖 AI Agents & Infrastructure
- Reason Machines continued promoting its AI-agent ecosystem. Its Reason Network lets coding agents such as Claude Code and Codex communicate across computers and exchange tasks. The company offers a free peer-to-peer network with an open-source communication engine.
💻 AI Coding & Developer Tools
- Anthropic confirmed new Claude Code restrictions against copying existing games. The company added instructions intended to prevent unauthorized recreation of specific games. Some developers also report that Opus 5.5 refuses legitimate reverse-engineering work, although Anthropic has not confirmed the full scope of those refusals.
🔬 AI Research & Models
- Isomorphic Labs president Max Jaderberg predicted that AI will eventually design all medicines. His comments describe a long-term ambition for AI-driven drug discovery. Clinical trials and regulatory approval would still be necessary before an AI-designed treatment could reach patients.
🏛️ AI Policy, Governance & Safety
- Legal researchers warned that AI could discover exploitable weaknesses in legislation. Scholars Felix Ringe and Linus Folkerts compared legal inconsistencies to software vulnerabilities, arguing that AI could search statutory language at a scale previously unavailable. Legal experts cautioned that human analysis would still be necessary to determine whether a suspected loophole has legal consequences.
- Emergency alert apps in Ukraine and Russia were hacked, resulting in false warnings about nuclear and other attacks. Both countries reported breaches, but the perpetrators and any connection between the incidents remain unconfirmed. The attacks show vulnerabilities in systems people rely on for urgent information; public reports have not established that AI was involved.
📊 Fundraising & Deals Roundup
- Anthropic's expanded infrastructure commitments drew scrutiny from financial analysts concerned about the scale of computing investments required for future AI development. The Australian Financial Review examined how those obligations could affect confidence in the industry's spending plans.
- Blackstone-backed QTS returned to the debt market with a proposed financing arrangement similar to one it withdrew earlier in the year. The transaction could demonstrate how investors are pricing the financial risks of AI infrastructure.
- Roleo raised €2.9 million to expand its AI career agent across Europe. The platform monitors opportunities on behalf of technology professionals and can introduce candidates to employers when relevant positions emerge.
🎙️ Interviews, Panels & Podcasts
- Y Combinator CEO Garry Tan said he spends up to $200,000 annually on AI tokens. During a conversation with Walter Isaacson, Tan described his extensive use of AI coding agents and put his annual spending at approximately $100,000 to $200,000. Some people collect sneakers. Apparently, others collect token bills.
Previous Around the Horn Digests
Catch up on recent issues:
- Monday, November 2: OpenAI changed ChatGPT and Codex usage plans, while Apple prepared its AI rollout in China.
- Friday, October 30: OpenAI revisited an agent escape, Reddit questioned AI-search traffic, and Mistral opened access to its trillion-parameter model.
- Thursday, October 29: Google restricted access to Gemini 4 Carbon while an OpenAI lawsuit raised questions about medical advice.
- Wednesday, October 28: OpenAI reportedly pursued Google chips, and Anthropic expanded government access to Claude.
- Tuesday, October 27: OpenAI reported new mathematical research, and researchers tested Anthropic's defenses.
- Monday, October 26: OpenAI described rogue-agent incidents, while Sweden restricted chatbots on government computers.
- Friday, October 23: AI agents reached banking workflows, and a supply-chain breach exposed new security risks.
That's a Wrap
That's more than 30 AI developments and product updates in today's roundup. If you made it all the way here, congratulations. You now know about phones that want to operate themselves and models that may eventually help create their successors.
For the shorter daily version, subscribe to The Neuron. We read the AI news so your browser doesn't need another 37 tabs open.
See you tomorrow.
P.S. Know someone trying to keep up with AI? Forward this roundup and invite them to subscribe.
