Everything That Happened in AI Today (Tuesday, October 20, 2026)

AI agents carried out cyberattacks while Microsoft introduced controls that can restrict or shut down risky agents.

Welcome, humans!

AI agents showed up on both sides of the security problem today. Reuters reported that a hacker used agentic tools to breach more than 100 company sites with little human intervention, while Microsoft introduced enterprise controls that can assess agent risk and stop actions. OpenAI, meanwhile, is trying to accelerate a massive funding round before Anthropic’s expected IPO.

The launch queue was busy too. New models competed on speed and cost, while several new tools gave agents more access to browsers or computers.

🆕 NEW From The Neuron

• I Tested Mercury 3. Diffusion Language Models Just Got Serious: Corey tested Inception’s new diffusion language model, which promises unusually fast text generation while competing with smaller frontier models on quality.
• Anthropic Consults Lawyers on How AI Agents Should Negotiate: Anthropic is exploring agents that can negotiate prices or disputes for users, which raises questions about how much authority people should delegate.
• Anthropic Is Still Hiring. Junior Technical Roles Are Taking a Smaller Share: Anthropic’s technical workforce kept expanding, but junior employees’ share of new hires fell from 27% to 9%.

Around the Horn — Tuesday, October 20, 2026

A hacker appears to have turned AI agents into a low-cost cyberattack team. Reuters reports that a Chinese-speaking hacker broke into websites belonging to more than 100 companies after assigning AI agents much of the work.

The agents handled complex parts of the intrusions with limited human involvement. Instead of manually directing every technical step, the attacker could give the system a goal and let the agents carry out much of the attack sequence themselves. The campaign reportedly cost only a few thousand dollars.

The economics are as important as the automation. A single attacker can now attempt more intrusions with less manual work, which changes how defenders have to think about scale. Security teams also have to watch employee-launched agents that receive legitimate system access, a problem Microsoft is now addressing from the other side.

🏆 TOP 5 NEWS (Around the Horn)

• Microsoft wants to stop rogue agents before they act. Microsoft introduced new security controls in Microsoft 365 E7 and Project Perception that assess individual agents for risk. Administrators can restrict an agent from making changes, require human approval, or shut it down.
• OpenAI is reportedly speeding up a huge funding round before Anthropic goes public. Reuters reports that OpenAI is racing to secure investor commitments in a round that could value the company above $1.4 trillion. The push comes ahead of Anthropic’s expected IPO and increases competition for investors willing to fund enormous AI infrastructure spending.
• DHS ordered DeepSeek removed from federal civilian systems. The Wall Street Journal reports that federal civilian agencies must remove DeepSeek technology. The order also blocks future purchases of the Chinese company’s products and services.
• Meta showed how it plans to supply computing power for increasingly large AI models. The company published an inside look at Prometheus, its gigawatt-scale AI supercluster in Ohio. Meta says the infrastructure has supported its Muse family of models as it expands toward much larger training and inference systems.
• Canadian news publishers sued Anthropic over copyrighted material. A coalition that includes Postmedia and The Globe and Mail filed a lawsuit in Ontario, alleging that Anthropic used publishers’ content to train Claude without permission. The claims remain allegations before the court.

Honorable Mentions

• The data center boom could accelerate sharply in 2027. SemiAnalysis expects roughly 50 GW of new capacity to arrive in one year, with another 60 GW already under construction. Its analysis suggests the next limiting factor could move from available power toward chips and memory.
• Some smart TV owners are using AI to remove the “smart” parts. TechCrunch reports that users are asking AI agents to strip ads or tracking software from TVs and replace sluggish interfaces. The experiments also show how coding agents can modify consumer devices that were never designed around agent-based customization.
• Claude Code reportedly found a crypto miner on an overheating Mac. A developer working on a temperature-monitoring app said the coding agent flagged a hidden cryptocurrency miner running on the computer. It is an anecdotal case, but it shows how coding agents can surface suspicious processes while investigating an unrelated technical problem.

🍪 TOP TREATS TO TRY

• Mercury 3: Inception CEO Stefano Ermon introduced Mercury 3, which the company calls its most capable diffusion language model yet. Inception claims more than 1,000 tokens per second and says the model outperforms GPT-6 Luna and Claude Haiku 5.5 on its selected quality comparisons. Diffusion language models generate and refine groups of tokens in parallel instead of producing text strictly one token at a time.
• MiniMax M3.1 Flash: MiniMax launched M3.1 Flash, with claimed gains in coding and computer use. The company says online inference is nearly three times faster while usage costs are 70% lower. OpenCode also revealed that the mystery “Space Bunny” model users had been testing was M3.1 Flash, while MiniMax shared its own launch announcement.
• ChatGPT browser actions: The official ChatGPT account announced expanded browser capabilities that let ChatGPT use open Chrome tabs as context. It can also take actions such as filling forms or downloading files through the browser integration.
• FLUX 3 Max: Black Forest Labs introduced FLUX 3 Max, a customizable video model that it says can generate a 10-second clip in about 12 seconds. The company also claims a 22× speed advantage over Seedance 2.5 at comparable reported quality.
• TorchNeuron: Amazon opened TorchNeuron in public beta, with plans to release it as open source. Developers can move PyTorch workloads to Amazon Trainium chips with a basic configuration change rather than rewriting them around Trainium-specific software.
• ChatGPT for older programming languages: OpenAI is improving ChatGPT’s coding capabilities for languages including COBOL and Fortran. The work adds evaluations aimed at software engineering tasks and COBOL migration, where modern AI coding tools have had less training material.
• SKY-001: Skyfall AI previewed SKY-001, which it describes as an “engineering world model” for tasks such as computer-aided design and building information modeling. The company reports substantial speed and cost advantages over general-purpose frontier models on its internal engineering evaluations.
• Drex Max: Nace AI launched Drex Max, a multimodal decision model designed for agents that work with text or visual interfaces. The company claims 34-millisecond decisions and pricing of $0.06 per million input tokens.
• Optimus-1: ArkenOS introduced Optimus-1, described as a swarm system for autonomous research. It can create multiple agents that explore different hypotheses across a research problem before combining the results.
• OpenClaw v2026.10.5: The latest OpenClaw release adds a sidebar for quicker agent access and Undo support in Skill Workshop 2.0. Public conversation links can now be opened without signing in.
• Anthropic OSS Scanner: ClaudeDevs launched OSS Scanner, which gives critical open-source projects free periodic vulnerability scans from Anthropic models. Findings can include a proof of concept and, when possible, a proposed patch.
• Sunra MCP: Sunra launched an MCP service that brings image and video models into assistants such as Claude or ChatGPT. Sunra says customers pay each model provider’s API price with no additional markup.
• Beam.win: Beam lets you keep agents or builds running on a Mac mini or Mac Studio while viewing their Terminal, browser, or app windows from another Mac. Remote windows appear like local apps with their own Dock presence.
• SpookyKit: LiveKit released five Halloween-themed voices for voice agents through custom Fish Audio voices. This one is mostly for fun, unless your customer-support bot has been asking for a seasonal career change.

💼 AI Productivity, Labor & Economics

• Six-foot inventory robots are becoming a familiar grocery-store sight. Yahoo reports that Tally robots are roaming stores such as Ralphs to scan shelves for missing products or misplaced items. They can also flag pricing discrepancies for employees.

🔬 AI Research & Models

• AI is entering more specialized clinical discussions. In an interview about myelofibrosis, researcher Adrián Mosquera Orgueira discussed AI’s possible role in medical decision-making and the importance of assessing systems across diverse real-world patient scenarios.

💡 Industry Commentary & Analysis

• Egen landed on CRN’s inaugural 2026 AI Innovators list. The technology services company announced the recognition, which highlights companies working on enterprise AI services and implementation.

Previous Around the Horn Digests

Catch up on everything you missed:

• Monday, October 19, 2026: OpenAI disclosed four rogue-agent incidents, Anthropic’s junior technical hiring share fell sharply, and OpenCode reported a breach tied to a model-distillation attack.
• Friday, October 16, 2026: Bank of America hired 1,000 contractors to patch flaws exposed by Anthropic’s Mythos. Eight states also sued OpenAI over ChatGPT safety claims.
• Thursday, October 15, 2026: Anthropic warned about small AI models that could replicate across internet-connected devices, while CoreWeave bond yields neared 15%. AI stocks also sold off as rate fears returned.
• Tuesday, October 13, 2026: OpenAI said its models helped investigate a hidden office recorder, and Google opened Gemini 4 Argon more widely. AI-authenticity disputes also spread into hiring and publishing.
• Friday, October 2, 2026: Frontier models beat a CPA baseline on month-end accounting tasks while Ataraxos defeated a Stratego champion. OpenAI also funded 163 nonprofits.
• Thursday, October 1, 2026: OpenAI parted ways with three safety researchers over alleged information sharing, while Anthropic reportedly moved toward a November IPO. New agent and coding tools also arrived.
• Wednesday, September 30, 2026: The FTC opened an AI-lab probe while DeepMind watermarked AI-designed proteins. ElevenLabs also reached a reported $22 billion valuation.

That's a Wrap

That’s 28 developments pulled from 30 links today. If you made it to the bottom, you survived a launch list long enough to need its own scrollbar.

For the daily version in a five-minute read, make sure you’re subscribed to The Neuron. We send six issues a week and sort through the firehose so your inbox doesn’t have to.

See you tomorrow.

P.S. Know someone who follows AI this closely? Forward this to them and tell them to subscribe here.