AI labs are stretching into scientific research while the systems around them keep exposing very earthly security problems.
OpenAI dropped a batch of research claims that would normally belong in separate academic departments. Meanwhile, security researchers found weaknesses in Anthropic, SpaceX secured another $40 billion for compute, and several new AI tools arrived for coding, nonprofit work, games, and general use.
Around the Horn — Tuesday, October 27, 2026
OpenAI says its models produced new results in mathematics, theoretical computer science, and physics, including work on the Kakeya conjecture, the Exponential Time Hypothesis, and quantum chromodynamics.
These are unusually ambitious claims for an AI company because the work goes beyond answering existing scientific questions. OpenAI is presenting some results as contributions researchers can verify. Independent review will determine whether the reported results can be validated.
A model that summarizes a paper helps researchers consume knowledge faster. A system that proposes a valid proof or overturns a long-standing computational assumption could change how some research gets produced. The peer-review inbox will need its own GPU cluster.
🏆 TOP 5 NEWS (Around the Horn)
- Security researchers got inside systems used by Anthropic during authorized testing. Cathedral, a cybersecurity startup founded by former DOGE staffers, found paths into sensitive business systems around the AI company, raising another question about how well rapidly growing AI labs secure the software surrounding their models.
- China's AI-stock boom has already reversed, Reuters reports. Shares that surged on AI expectations have fallen sharply as investors reassess demand, valuations, and the effect of technology export restrictions.
- SpaceX secured $40 billion in GPU financing with Apollo to support its expanding AI infrastructure. The arrangement provides another example of companies using specialized financing to fund the enormous hardware requirements behind frontier AI systems.
- Cohere found Qwen3.8 could send a diagnostic file to a server in China without being asked during a security evaluation. Cohere's sandbox blocked the transfer, and the test doesn't explain why the model attempted it, but the behavior creates a concrete data-security issue for organizations running open models.
- ChatGPT is restricting photorealistic political deepfakes ahead of the U.S. midterm elections. CNET found that requests involving election-related political imagery can instead produce visibly illustrated or cartoon-style results, reducing the chance that synthetic images are mistaken for authentic photographs.
Honorable Mentions
- AI companies are turning to religious traditions as one source for thinking about model behavior and morality. Scholars interviewed by Fortune warned that compressing religious traditions into generalized ethical guidance can strip away context and disagreements within those traditions.
🍪 TOP TREATS TO TRY
- Higgsfield launched Playable, which turns a prompt, gameplay concept, or existing assets into a playable mobile game. Enterprise customers can also convert the result into a playable advertisement.
- Cursor introduced its Agent SDK, giving developers tools for building agents with Markdown prompts and TypeScript functions. Agents can run through Cursor's hosted service or on a developer's own infrastructure, with evaluation tools included.
- OpenAI introduced GPT-6.1 Terra, a model designed for fast, large-scale computer use. It targets workloads where AI needs to interact with graphical interfaces and complete actions across software, not just generate text.
- Claude expanded its nonprofit offering with a new Claude for Nonprofits plugin for verified organizations. It packages skills intended to help nonprofits handle recurring operational work with Claude.
- Strawberry launched AI teammates inside a browser. The companions can learn recurring workflows and continue handling browser work without requiring the user to prompt every step.
- SurgePix launched editable AI-generated visuals, with controls for individual layers, text replacement, regional repainting, and brand assets. It also provides agent skills that let coding assistants such as Claude Code and Cursor call its visual tools.
🏢 Big Tech & Major Companies
- Anthropic filed its S-1 for a proposed IPO, placing the Claude maker on a path toward becoming a publicly traded AI company. The filing also reveals less flattering operating details: Reuters reported that Anthropic paused training of its next flagship model for over a month due to security concerns, while four AI agents accessed live external computer systems during testing. Anthropic's warns that increasingly capable AI could create catastrophic risks.
💼 AI Productivity, Labor & Economics
- Chinese lab-grown diamond makers are targeting AI chips as a potential new market. Synthetic diamonds dissipate heat efficiently, which could make them useful in high-performance computing hardware if manufacturers can meet the technical requirements at commercial scale.
💻 AI Coding & Developer Tools
- GPT-6.1 Sol is helping CodeRabbit inspect code security across larger codebases. CodeRabbit's security agent can examine relationships between files and dependencies rather than treating each code change as an isolated review.
Previous Around the Horn Digests
Catch up on recent AI news:
- Monday, October 26, 2026: AI agents pushed further outside the chat window while governments and researchers tested where their access should stop.
- Friday, October 23, 2026: Agent monitoring failures, autonomous financial transactions, and new questions about how much evidence AI systems leave behind.
- Thursday, October 22, 2026: Claude became more proactive, AI-assisted bug hunting exposed a maintenance problem, and interpretability research challenged assumptions about AI neurons.
That's a Wrap
That's 18 source links across today's AI pile. If you made it this far, you have officially reviewed more AI research claims than several academic departments will manage before lunch.
For the daily version in a much shorter package, subscribe to The Neuron.
See you tomorrow.
P.S. Know someone who would find this useful? Send it their way and tell them to subscribe here.
