Subscribe Forward this edition

The Agentic Enterprise.

AK · Daily Edition · 8 min read

Wednesday, November 4, 2026

AI agent killswitch needs a rulebook

CrowdStrike shipped a kill switch for AI agents and agreed to buy two agent-security startups, the same day Apple warned investors that agents may act without permission.

Two documents landed on Tuesday that read like halves of one plan. Apple's annual report listed what can go wrong when agents act on their own. CrowdStrike sold tools to catch it: a switch to stop an agent and a recorder for what it did. In most companies those halves belong to different teams. Security will own the switch. Someone still has to write down when to pull it.

The Big Story

Governance

Security can stop an agent. Governance has to decide which ones.

C

rowdStrike's AI Kill Switch, now live in Falcon Guardian, lets an authorized administrator stop a rogue agent, block it from relaunching and cut its network access, with every action logged. CrowdStrike is also buying Fiddler AI, which records what agents do and why, and Dreadnode, which attacks models and agents before they ship. The WSJ put the price at about $495 million, per TechCrunch; CrowdStrike didn't disclose prices. The Next Web · TechCrunch

CrowdStrike got there second. In May, ServiceNow added what it called a kill switch to its AI Control Tower, which shuts down an agent that goes beyond its permissions, inside a product that also scores agent risk against NIST and EU AI Act frameworks. A security vendor and a workflow vendor have built the same button from opposite ends: one from the security operations center, one from the governance register. ServiceNow

Both switches line up with the risks in Apple's new 10-K. Apple warns agents "may behave in unintended or unpredictable ways," may act on systems they aren't authorized to reach, and may "conceal actions, hide, withhold or manipulate information." A kill switch answers the first two risks and Fiddler's recorder the third. Dreadnode maps findings to OWASP, MITRE ATLAS and NIST's AI framework. AI Weekly · The Next Web

What no vendor sells is the decision about when to press it. CrowdStrike's switch works at the gateway today, with endpoint blocking coming "soon," and Apple names the harder limit: deleted data or a completed transaction may be "difficult or impossible to reverse." Past that point, stopping the agent is cleanup. Governance has to act earlier, deciding which actions need approval and what the recorder must keep.

A kill switch without an owner is a button nobody is allowed to press.

The Spearhead Take

Give every production agent two named owners: a governance owner who sets what it may do without approval, and a security owner who can stop it. Write down its irreversible actions first, then test the switch against those.

Sources: The Next Web · TechCrunch · ServiceNow · AI Weekly

The Obvious & The Overlooked

What is easy to see, and what is easy to miss.

The Obvious

Contact centers are where agents pay first. Salesforce agreed to buy Cresta, its second reported multibillion-dollar customer service deal this year. Source: SiliconANGLE.

Chip demand hasn't cooled. AMD's data center revenue nearly doubled and it guided to about $17 billion next quarter. Source: AMD.

Agents are moving into the operating system. Apple now lets developers wire app actions into Siri through App Intents in the iOS 27.3 betas. Source: Apple.

The Overlooked

Power is becoming a negotiating chip for data centers. OpenAI is offering utilities the right to cut its data center load during demand peaks. Source: tEDmag.

You may not be able to prove which model answered you. Anthropic published research code to verify a response came from the model claimed. Source: TechCrunch.

The model router does the unglamorous work in a cost cut. Revolut's self-hosted model handles code while a router sends what it can't do elsewhere. Source: The Register.

Release design sets how fast models improve. Meta ships ads features faster by letting each one deploy and roll back alone. Source: Meta Engineering.

Moving Pieces

Five developments with practical consequences.

Deals

Salesforce buys Cresta, and the voice agent goes in-house

Salesforce agreed to buy Cresta, which builds voice agents for customer service and real-time coaching for human reps. Salesforce didn't disclose terms; The Information put the price at $2.7 billion. It follows the $3.6 billion Fin (formerly Intercom) deal and is expected to close in Salesforce's fiscal 2028, which begins in February. If you run Cresta beside another CRM, start the neutrality conversation now. Roadmaps tend to bend toward the new owner's platform within a year or two. SiliconANGLE · The Next Web

Markets

AMD's data center business nearly doubled

AMD reported third-quarter revenue of $13.4 billion, up 45%, with data center revenue of $8.5 billion, up 97%. It guided to about $17 billion for the fourth quarter. AMD says Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud are running workloads on its first Helios rack systems built on MI450 GPUs, and it booked a $1.1 billion non-cash charge on warrants issued to customers. A credible second GPU supplier gives buyers leverage on price and delivery dates, but only if your software can run on both. AMD

Deployment

Revolut ran its own model and cut coding costs about 90%

Revolut replaced Claude Code with a self-hosted open-weights model, Z.ai's GLM, as the default coding assistant for about 2,000 engineers, and cut AI coding inference costs roughly 90% in under a year, its infrastructure lead told an AI Engineer conference, per The Register. The setup runs on 56 Nvidia B200 GPUs. A router still sends vision queries to Gemini or GPT. The biggest gains came from applying it to debugging and ticket work. The Register

Engineering

Meta ships ads model features 3.6 times faster by letting each one fail alone

Meta rebuilt F3, the system that feeds features to its ads ranking models, so each feature compiles, deploys and rolls back on its own instead of inside one merged release. One ranking request can need more than 20,000 features. Meta says comparable features now reach production about 3.6 times faster with no added compute. For your own ML platform, the lesson is about release design. One bad feature shouldn't hold up the other 19,999. Meta Engineering

Research

Anthropic wants answers you can check against the model's fingerprint

Anthropic released FlashProve, open-source research code for "provable inference": a lab publishes a fingerprint for each model, and anyone can cryptographically check that an output came from it without seeing the weights. The use case is simple. Did your vendor swap in a cheaper model to save money? Anthropic calls it intermediate progress and says no lab, itself included, plans to deploy it yet. Keep it on the list for model-change clauses in contracts. TechCrunch · Anthropic

On the Radar

Eight signals worth keeping in view.

Deployment

ADNOC signed a multi-year deal with OpenAI and G42 to deploy OpenAI Frontier across its operations. ADNOC calls itself the first energy company to do so, alongside more than 200 existing AI applications. ADNOC

Product

OpenAI says ChatGPT's web searches equal roughly a fifth of Google's query volume. The figure uses OpenAI's own definition; the company also cited 1.2 billion weekly users. Big Technology

Product

Soniox shipped STT v6, speech recognition in 60+ languages built for voice agents. It filters out the agent's own audio while still catching a caller who interrupts. Soniox

Research

OpenAI says two employees used GPT-6 Astra to design a working hardware prototype in under a week. Several revisions were needed because the model missed manufacturing tolerances, The Decoder reports. The Decoder

Policy

AI-generated political ads became a midterm issue, with researchers tallying nearly $187 million in spending, CNBC reports. Disclosure rules still vary state by state. CNBC

Workforce

Google DeepMind staff challenged contract language that appeared to limit criticism after they leave, the Washington Post reports. Google says it doesn't enforce it and removed it from new contracts. Washington Post

Research

AI agents now run operations at SETI's Allen Telescope Array. Some analysis tasks fell from weeks to hours, and the system spotted a known pulsar in real time. The Decoder

Research

Drugmaker CEOs told the WSJ that AI helps find targets but biology remains the bottleneck. No AI-designed medicine has FDA approval yet. WSJ

Quick Hits

The wider field, one line each.

Raises

01Wabi raised $55M at a $1.25B valuation, led by IVP, for a personal AI built for groups.Ventureburn
02Opentrons raised a $40M Series E for lab robots that run AI-driven experiments.GEN
03NexGen Cloud, a UK AI infrastructure provider, raised $106M.Parsers VC
04TrashLab raised $12M for an AI operating system for waste haulers.Parsers VC
05Querio raised $5M from BMW i Ventures and Haatch for AI business analytics.Parsers VC

Ships

06Scalekit launched AgentKit, one gateway to more than 30,000 tools across 600 apps.Product Hunt

People and research

07OpenAI named former Uber executive Jill Hazelbaker chief corporate affairs officer.Axios
08Cat Wu, who led product for Claude Code, has left Anthropic.Business Insider
09VagueBench found embodied agents' task success fell about 20% when background context was removed.arXiv
10El Agente Fármaco, a drug-discovery agent, beat plain tool calls on a 22-task benchmark using fewer tokens.arXiv

From Spearhead · AI Bootcamp

Get your leadership team AI-ready, in a day.

The AI Bootcamp is a one-day, in-person session designed for your industry, your organization, and your use cases. Your leaders understand AI, build AI agents, envision outcomes, and start building a new work culture.

Plan a Bootcamp for your team

The Number

1,000 MW

The flexible demand OpenAI's planned Georgia campus could shed when the grid is strained, under its new Grid Protectors program.

OpenAI likens it to the average demand of about 700,000 households, and says it will pay its own energy and infrastructure costs. The campus is still in development. Expect utilities to ask large buyers, colocation tenants included, for similar terms. tEDmag

Counter-Signal

Security

The labs are still fixing the basics

Agent controls sit on top of ordinary security, and this week showed that layer under test even at the frontier. OpenAI said it is working with Cathedral, a startup founded by former DOGE staffers, to test its defenses, and is more than quadrupling rewards for critical bug-bounty findings. Cathedral reportedly reached parts of Anthropic's internal systems earlier through Anthropic's bug-bounty program; it told the Washington Post it got to Slack messages and meeting transcripts in under 12 hours. Anthropic said it took several days, relied on multiple third-party vulnerabilities, was out of scope, and said its most sensitive systems, including model weights and Claude conversations, were not affected. Gizmodo

By Anthropic's account, the path ran through third-party software vulnerabilities. An agent kill switch won't help much if the systems around the agent leak first.

From the Field

In most companies, the people who write AI policy and the people who run the security operations center have rarely sat in the same meeting.

That made sense when AI meant a chatbot answering questions. It makes less sense once an agent can issue a refund or email a vendor. Governance teams write rules about what agents may do. Security teams build the means to stop them. Each side holds half of the answer to which actions matter most.

In client work, the most useful starting point has been one shared list: the actions an agent can take that can't be undone. Payments, deletes and outbound email tend to make it. Drafts and queued changes usually don't. That list tells governance where approval belongs, and it tells security where the switch has to work first, on the endpoint as well as at the gateway.

At most companies the list is short. It's also usually missing, and the tools in this week's deals have nothing to aim at until it exists.

Let's create value with AI,
AK

Work with Spearhead

AI Bootcamp

One day to get leaders fluent enough to decide.

AI Governance

Guardrails up front, controls in the build.

AI Engineering

Production systems your team owns on Monday.

The Agentic Enterprise

Know more about AI than 95% of your peers. Every morning.

OpenAI, Anthropic, NVIDIA, AWS and Azure are Spearhead technology partners (Microsoft through Azure); CrowdStrike, ServiceNow, Apple, AMD and Salesforce are not. Produced using Anthropic’s Claude with editorial direction from AK.

© 2026 Spearhead. All rights reserved.