Everything That Happened in AI Today (Friday, October 23, 2026)

AI agents are getting access to systems where a wrong action can move money, expose corporate accounts, or affect a medical decision.

Welcome, humans!

Today’s stories show autonomous AI moving into environments with consequences beyond a bad chatbot answer. Banks are trying to understand third-party agent activity inside customer accounts, while security researchers are testing whether AI companies can reliably monitor their own autonomous systems.

Research moved quickly too. Anthropic used Claude for software verification at scale, while Microsoft researchers reported progress in medicine and a 31-year-old mathematics problem. A separate release wave brought new tools for vector graphics, video creation, agent credentials, and scientific research.

🆕 NEW From The Neuron

• Claude Duet Can Prepare Your Work. How Much Will You Need to Check?
  Claude Duet prepares work across connected apps, but its value depends on whether the time it saves exceeds the effort users spend checking its recommendations.
• PHP2Shell Exposes the Maintenance Challenge Behind AI Bug Hunting
  AI-assisted security research can uncover vulnerabilities faster than software maintainers can repair them, creating a new challenge for widely used software projects.
• Could AI Models Be Easier to Understand Than Scientists Thought?
  Goodfire found evidence that individual AI neurons may encode more coherent functions than researchers previously assumed, which could change how scientists study model behavior.

Around the Horn: Friday, October 23, 2026

AI agents have started showing up inside U.S. bank accounts, and banks are having trouble determining exactly what those agents are authorized to do.

Axios reports that major banks, including Morgan Stanley and Bank of America, are racing to contain risks from unvetted third-party AI agents that can access customer accounts. The central problem is attribution: a bank needs to know whether an action reflects the customer’s instructions or an agent operating beyond them.

Agent access to financial accounts turns permission design into a financial control issue. An agent that reads balances presents one type of risk. An agent that can initiate transactions or alter account settings gives banks a much harder authentication problem, especially as consumer AI products gain more autonomy.

The banking industry now has to decide what an AI agent can do on a customer’s behalf and how institutions can verify the customer authorized each consequential action.

🏆 TOP 5 NEWS (Around the Horn)

1. China-linked hackers used agentic AI to accelerate a major supply-chain attack. Microsoft said Storm-0940 used the ARTEX agentic penetration-testing system during a ConnectWise compromise. The agents automated attack activity and abused application credentials across Microsoft 365 environments, showing how autonomous tools can compress offensive cyber operations.
2. Microsoft says its new medical AI system beat general physicians on difficult cases. The system works through diagnoses sequentially rather than producing a single immediate answer. In a study published in Science, it was more than twice as accurate as general physicians on complex cases. TechCrunch noted that much of the improvement may come from the underlying AI models rather than Microsoft’s diagnostic system itself.
3. External evaluators found vulnerabilities in OpenAI’s internal agent-monitoring system. METR and Redwood Research spent three weeks embedded at OpenAI to evaluate systems designed to detect misalignment in internal agents. Evaluators found several vulnerabilities, including one that could have let agents hide hundreds of actions from monitoring.
4. Anthropic used Claude to verify security properties across 123 open-source libraries. The company says large language models are making software verification feasible at a much larger scale. Claude proved properties about compiled code and uncovered bugs that had escaped detection for years. The verification uses mathematical proofs to determine whether software behaves according to specified rules.
5. Microsoft researchers resolved a 31-year-old mathematics problem with help from human-directed AI research. Xinzhi Zhang and collaborator Zhao Song proved the Kannan-Lovász-Simonovits conjecture, a major problem in high-dimensional geometry and probability. The work used AI to explore candidate mathematical directions while researchers selected and refined the approaches.

Honorable Mentions

• Sam Altman says AI-discovered security flaws may exceed the assumptions behind conventional patching. The OpenAI CEO said models can find vulnerabilities outside normal hacker threat models, including hardware-level problems, and questioned whether patching every newly discovered weakness will remain feasible.
• Bill Gates expects even a favorable AI transition to produce two turbulent decades. Gates described AI as an extraordinary source of potential progress while warning that the technology could create major disruption and severe security risks.
• OpenAI fired three safety researchers over sensitive Signal conversations. The researchers were dismissed for mishandling sensitive information amid an internal dispute over how much access outside researchers should have to AI systems.

🍪 TOP TREATS TO TRY

• Claude is making auto mode the default for more paid users. Starting next week, sessions on Claude’s web, desktop, and mobile apps will use auto mode by default for Pro, Max, and Team plans. Automated checks will block risky actions, while users can return to manual approvals.
• 1Password introduced a platform for AI agents. 1Password for Agents is being developed with companies including Anthropic and OpenAI. The platform offers autonomous agents access to credentials and services through 1Password’s security system.
• Ideogram can now generate editable vector graphics. Its first vector model produces scalable SVG files from text prompts and can convert existing images into vectors, with support for typography and gradients.
• DataVendor opened the public beta of its AI-data marketplace. The platform is designed for researchers and companies that need to buy or sell datasets used for AI development.
• HeyGen released its first general-purpose video model. HeyGen Video 1 ranked first for text-to-video in a blind Artificial Analysis test.
• Canva launched its first image-generation model. Canva is making the model available to all users and is pitching it as a more customizable, lower-cost option for image creation inside its design platform.
• OpenAI expanded access to its decision model. OpenAI is making the model broadly available as competition grows around AI systems built specifically for decision-making tasks.
• Hugging Face released Carbon-X for DNA research. The model can analyze DNA sequences and generate new ones, giving scientists a natural-language interface for genomic research. Potential applications include gene-therapy research.

🏢 Big Tech & Major Companies

• Arc is shutting down as The Browser Company concentrates on AI browser Dia. Arc will become inaccessible on December 22, ending the browser that helped establish The Browser Company while the company shifts its resources toward its AI-centered successor, Dia.

🤖 AI Agents & Infrastructure

• Anthropic is donating Agent Skills to the Agentic AI Foundation. Claude Developers said more than 70 agent platforms have integrated the standard since December 2025, including ChatGPT, Cursor, GitHub Copilot, and VS Code. Anthropic previously donated Model Context Protocol to the same foundation.

📊 Fundraising & Deals Roundup

• OpenAI is reportedly working on another $30 billion fundraising round. The round could value OpenAI at roughly $1.25 trillion ahead of a potential IPO, which would put the company above the valuation of any previous private company if completed on those terms.

Previous Around the Horn Digests

Catch up on the previous AI news cycles:

• Thursday, October 22, 2026: AI-discovered PHP vulnerabilities triggered emergency fixes while Google DeepMind faced resignations over military AI work.
• Wednesday, October 21, 2026: OpenAI acquired Orchestra Bio, while Anthropic proposed a federal AI safety bill.
• Tuesday, October 20, 2026: An AI-assisted hacker breached more than 100 company sites as Microsoft introduced new controls for enterprise agents.
• Monday, October 19, 2026: OpenAI disclosed new rogue-agent incidents, while OpenCode reported a breach connected to model distillation.
• Friday, October 16, 2026: Bank of America brought in contractors to repair flaws exposed by Anthropic’s Mythos as states challenged OpenAI over safety claims.
• Thursday, October 15, 2026: Anthropic warned that small AI models could soon replicate across connected devices, while AI infrastructure financing showed new signs of strain.
• Tuesday, October 13, 2026: OpenAI said its models helped identify an employee accused of planting a hidden recorder, while Google expanded access to Gemini 4 Argon.

That's a Wrap

That’s 20 story clusters from 23 unique source links today. If you made it all the way down here, you have officially done more manual checking than Claude’s new auto mode would prefer.

For the daily version in a five-minute read, subscribe to The Neuron.

See you tomorrow.

P.S. Know someone who would want the full AI news firehose? Forward this to them and tell them to subscribe here.