Watch on YouTube · subscribe for the daily brief

From early 2027, Google's AI tools for K-12 students will be off unless a school turns them on. The catch: Google Search counts as one of those AI tools. A school that restricts AI altogether loses Search, and students see a message that it is not available. Google's education VP Chris Phillips told the New York Times that "Search and AI are deeply integrated."

What schools actually get to choose

• Off by default: Gemini, Gemini Notebook, Gemini in Classroom and Google Search, for student accounts.
• If Search is on: either the standard AI Overviews and AI Mode, or a new Guided Learning mode meant to walk students through problems rather than hand over answers (Ars Technica).
• Not offered: Search without AI. Los Angeles had asked for exactly that (The Decoder). New York City and Los Angeles both pulled back on AI for students in September.
• The backlash: Bernie Sanders called it "extortion"; New York mayor Zohran Mamdani said Google is using "its monopoly power to force AI into our public schools"; Lina Khan called it "the danger of allowing an unregulated monopolist to control an essential service." Google replied on X that "schools are in control", without addressing the Search point (CNBC).

For builders: if you sell into schools or run their IT, the 2027 switch is a procurement decision, not a settings toggle. Other engines, such as DuckDuckGo, let you turn AI off.

A kill switch for agents

CrowdStrike launched an AI kill switch in its Falcon Guardian platform. It stops an internal agent, blocks it from relaunching and cuts its network access. It does not delete the agent's code; it cuts the agent off from its models through connected AI gateways, and a human has to make the call. CrowdStrike is also paying around $400 million for Fiddler AI and about $95 million for Dreadnode, per the Wall Street Journal: one watches agents in production, the other attack-tests them before launch.

Agents on the attack: Chainstack and WorkOS

Chainstack disclosed a breach: in through a third-party component into a dev environment on October 26, then into parts of production, detected and cut off October 31. Exposed: organization and account holder names, member emails, password hashes. Systems with node credentials were reachable; Chainstack has "no evidence they were read" but asks customers to rotate them, regenerate API keys and change reused passwords. The activity overlaps with the China-linked ARTEX campaign, an AI agent framework for offensive operations. WorkOS reported an incident with the same framework, limited to its go-to-market systems.

Patch now: Siemens SIMATIC IoT2050

CVE-2026-58115, CVSS 10.0: with Node-RED installed, the HTTP interface needs no authentication, so anyone who can reach it can run code as root. VulnCheck sees complete exploits deploying a bot since October 30. Update to Industrial OS 4.3.4.1 or later, or uninstall Node-RED.

Also from the security desk: LLMjacking

Google's Mandiant and Threat Intelligence Group published a report on LLMjacking: attackers using stolen cloud credentials and AI keys to run up charges above $100,000 in a day, up to $1 million at some victims, and reselling the access for about $60 a month. Our full video breaks it down; the short version is to cap spend on every AI key and alert on any AI service you did not enable.

OpenAI's week

• A device in under a week: OpenAI says GPT-6 Astra designed a hardware device for talking to Codex, enclosure, PCB and firmware, as a side project in under a week. Several revision rounds were needed because it did not always account for manufacturing constraints.
• Bug bounty x4: critical bugs in ChatGPT or Codex now pay up to $30,000, and with bonuses up to about $172,000. Reports from the last 30 days get the difference paid retroactively.

Proving which model answered

Anthropic is working on FlashProve, an open-source system for "provable inference": a cryptographic proof that an output came from a specific model, meant to stop a provider from quietly swapping in a cheaper model.

Chips, power and the cloud

• Huawei: is telling some customers to source HBM memory for its AI chips themselves. Smuggled HBM3E went from a few hundred dollars to around $4,000 a unit in a year; a chip needs six to eight.
• Nvidia: has told some cloud customers to take on less business from Chinese companies, closing the remote-access gap in export rules. Not mandatory, future contracts only.
• Jet engines: some AI data centers now make their own power with converted jet engines burning gas: about 24 months to run, against about seven years for a 100 MW grid hookup in Northern Virginia.

Quick hits

• Apple will require an 18+ rating for any app with dating or romantic features, AI companions included, with new content descriptors from November 16.
• At the SETI Institute's Allen Telescope Array, AI agents can now steer the telescope and analyze radio signals as they arrive, on Nvidia Holoscan.

The through-line

Control. Google decides what schools can switch off, CrowdStrike sells a switch for agents, and attackers run agents of their own. One thing to do today: list which of your tools can turn their AI off, and which cannot.

Sources covered

1 story · 110 sources (55 news, 51 company, 4 voices)

1. How We Optimized Browser Use for Deep Agents
   🏢LangChain↗🏢LangChain↗🏢Confluent↗🏢Databricks↗🏢Databricks↗🏢Snowflake↗🏢Snowflake↗🏢Snowflake↗🏢MongoDB↗🏢MongoDB↗🏢MongoDB↗🏢MongoDB↗🏢MongoDB↗🏢Vercel AI↗🏢ClickHouse↗🏢Diraq↗📰The New Stack↗🏢Vercel AI↗🏢Google Cloud AI↗📰The New Stack↗📰TechCrunch↗📰The Hackers News↗📰TechCrunch↗📰CNBC↗📰BBC News↗📰TechCrunch↗🏢Hugging Face↗🏢Microsoft Research↗🏢Google DeepMind↗🏢Meta AI↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢AWS AI/ML↗📰TechCrunch↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Hugging Face↗🎙️Simon Willison↗📰BleepingComputer↗📰SiliconANGLE↗📰TechCrunch↗📰The Verge↗🏢NVIDIA↗📰Ars Technica↗📰The Verge↗📰The Decoder↗🏢ClickHouse↗📰The Hackers News↗📰Ars Technica↗📰TechCrunch↗📰The Decoder↗📰The New Stack↗🎙️Ben Lorica↗📰InfoQ↗🏢Snowflake↗🏢Snowflake↗📰BleepingComputer↗📰TechCrunch↗📰The Verge↗📰The Verge↗📰The Verge↗📰The Verge↗📰Ars Technica↗📰TechCrunch↗📰The Verge↗🏢OpenAI↗📰The Decoder↗📰TechCrunch↗📰LWN.net↗📰The Verge↗📰TechCrunch↗📰The Verge↗📰The Decoder↗📰The Verge↗📰BleepingComputer↗📰TechCrunch↗📰SiliconANGLE↗📰TechCrunch↗📰BleepingComputer↗📰The New Stack↗🏢Estuary↗📰SiliconANGLE↗📰SiliconANGLE↗🏢Google Cloud AI↗🏢OpenAI↗🏢OpenAI↗📰The Decoder↗📰The Decoder↗🏢NVIDIA↗🏢Supabase↗📰The Decoder↗📰InfoQ↗📰BleepingComputer↗📰MIT Technology Review↗📰The Decoder↗🎙️Simon Willison↗🎙️Simon Willison↗📰The Decoder↗🏢Vercel AI↗🏢Vercel AI↗🏢Temporal↗🏢PlanetScale↗🏢Redpanda↗🏢Replicate↗🏢DuckDB↗🏢MotherDuck↗🏢Vercel AI↗🏢Hugging Face↗🏢OpenAI↗🏢Hugging Face↗