Subscribe Forward this edition

Your security vendor wants to run your AI models | The Agentic EnterprisePalo Alto will pay $4.5B to train and host your models. Your security vendor wants more of the stack.

The Agentic Enterprise.

AK · Friday, November 20, 2026 · 7 min read

Friday, November 20, 2026

Your security vendor wants to run your AI models

Palo Alto Networks agreed Wednesday to buy Applied Compute for $4.5 billion, so it can train, host and secure open-weight AI models for enterprises.

Read it next to CrowdStrike's agent-security deal earlier this month and a run of funded startups selling the same idea, and the week has a shape. The vendors that already sit on your firewall and your identity layer now want the models too. That's convenient. It also ties more of your AI stack to one supplier's pricing and its bad days, and that's a trade a CIO should make on purpose.

The Big Story

Deals

Palo Alto wants to build the vault it guards

P

alo Alto Networks said Wednesday it will buy Applied Compute, which helps companies post-train and run open-weight models, in an all-stock deal worth $4.5 billion. The news came with its fiscal first-quarter results. Applied Compute's CEO becomes co-chief AI officer, and the deal should close in the February-to-April quarter. Palo Alto Networks via SEC · CSO Online

The pitch is sovereignty. CEO Nikesh Arora told analysts that open-weight models went from under 5% of token traffic to more than 60% in a year, a figure he didn't source. Palo Alto will "train and serve secure sovereign implementations" of those models, in the public cloud or inside a customer's own walls. It plans to buy GPUs and work toward more than 100 megawatts over three to five years, while still renting from cloud partners. A firewall company is turning, in part, into a compute operator.

Arora's own case for the deal deserves a slower read. He argues that deterministic controls "cannot anticipate every scenario that an agent is going to conduct," so enterprises "will have to train a supervisory AI model at the same time" as the application it watches. He also says Palo Alto can't intercept and understand somebody else's application. Follow that through and one vendor builds and hosts the model, then supplies its supervisor. Greyhound Research's Sanchit Vir Gogia called that kind of supervisor "a colleague, not a check."

Add the $25 billion CyberArk purchase, which brought identity, and one supplier now reaches from the firewall to the login, with models next.

Palo Alto will "train and serve secure sovereign implementations" of open-weight models.

The Spearhead Take

Keep the builder and the grader apart. If one vendor offers to train your model and sell you the supervisor that checks it, have someone else run the check. Get a costed exit in writing before model work moves onto a bundled platform; it's cheapest to negotiate while the vendor is still courting you.

Sources: Palo Alto Networks via SEC · CSO Online

The Obvious & The Overlooked

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

The Obvious

Security vendors want the agent layer. Palo Alto and CrowdStrike both made agent-related acquisitions within three weeks. Source: CSO Online.

Public markets will fund a frontier lab. Anthropic's first-day gain shows investors will still pay for a frontier lab, though the stock closed well below its open. Source: Forbes.

Agent-security money keeps coming. Bay, WhiteMatter, Metano and Disseqt raised about $90 million combined this month to police agents. Source: The Next Web.

The Overlooked

Owning GPUs puts a firewall vendor in the data-center business. It also raises a plain question: where will your data and models physically sit? Source: CSO Online.

Some agent failures happen between customers. The reported Muse mix-up is a data-isolation failure that a kill switch on one agent wouldn't catch. Source: TechCrunch.

Federal agent rules tend to travel. Controls vendors build for agencies, like agent inventories and access caps, usually get sold to everyone next. Source: House Homeland Security Committee.

Moving Pieces

Infrastructure

AWS gives agents a shopping cart

AWS said November 19 that AI agents can now find, buy and deploy AWS Marketplace software through Bedrock AgentCore payments, billed to the customer's AWS account. Buyers can set spend limits, require approvers by dollar amount or product type, and route purchases through Coupa or SAP Ariba. CloudTrail logs every agent transaction. Each buyer's own policy decides which purchases need a person's sign-off; for a small sign-in tool, AWS's Matt Yanchyshyn told Fortune, "You shouldn't have to approve that."

Sources: AWS · Fortune

Deals

Workday buys a rulebook for agent math

Workday said November 18 it will acquire Honeydew, a semantic layer that turns a company's metric definitions into governed queries before a request reaches the model, and checks permissions at query time. Terms weren't disclosed, and the deal should close by the end of January. "An AI model shouldn't be improvising the calculation behind a payroll number," said Workday CTO Gabe Monroy. Most companies haven't written down what revenue or headcount means. Agents will find that gap fast.

Sources: Workday

Governance

Microsoft gets 40 companies to agree agents should keep receipts

At Ignite on November 17, Microsoft launched an AI Transparency Initiative with more than 40 organizations, among them Accenture, KPMG and Mastercard. The group will push open standards for how agents record identity, intent, model use and actions across providers. The idea is that an agent shouldn't be the only witness to what it did. No standard has shipped yet. Large buyers on that list want an audit trail they can read without taking a vendor's word.

Sources: PR Newswire · Channel Insider

Product

OpenAI puts a sales agent inside ChatGPT

OpenAI launched ChatGPT for Sales on November 17: a Sales plugin plus more than 30 partner plugins, including Salesforce, HubSpot and Gong. Reps can research accounts, prep calls and draft follow-ups from one chat window, and Business and Enterprise admins switch it on in the Admin Console. That moves ChatGPT into the revenue system, where companies keep some of their most guarded data. OpenAI publishes no independent performance numbers. The open question for IT is how much CRM data each plugin can read.

Sources: OpenAI

Policy

House bills would make agencies count their agents

House Homeland Security Republicans announced four AI-security bills on November 18. The AI INSIGHT Act would require federal civilian agencies to inventory the agents on their systems, limit each to the access its task needs, and revoke it when needed. The bipartisan ROOT CAUSE Act would set up a CISA board to investigate significant cyber incidents, AI failures included, with no enforcement power. None of the bills regulates private companies, though that board could review incidents on private systems. The INSIGHT checklist still reads like a sound first agent policy for anyone.

Sources: House Homeland Security Committee · Nextgov

On the Radar

Deals

Anthropic closed its first day at $97.17, up 17% on its $83 IPO price, after opening at $111.50; Forbes put its market value near $1.27 trillion. Forbes

Product

Google released Gemini 4 Neon for complex coding and enterprise workflows, with a 1-million-token context window priced at $3 and $15 per million input and output tokens. Google

Security

Meta's Muse agent reportedly mixed up conversations across accounts and exposed private data, including bank balances, according to a Reuters report covered by TechCrunch; Meta didn't immediately comment. TechCrunch

Policy

Citi and PNC changed their terms so customers answer for their AI agents' actions, and Citi's terms reportedly hold customers liable even after a hack. TechCrunch

Deployment

Salesforce in Microsoft Copilot entered public preview, bringing Salesforce customer context into Copilot as part of a pairing the two call Copilotforce, with Accenture as "Customer Zero." Salesforce

Product

Cloudflare open-sourced Kitesurf, its browser engine for AI agents, under the Apache 2.0 license, so teams can inspect the code their agents browse with. Cloudflare

Compute

Intel says it signed its first outside customer for its 18A-P chip process, unnamed, and named Microsoft the first customer for its EMIB packaging, as hyperscalers look for packaging capacity beyond TSMC. TrendForce

Workforce

Law firm HSF Kramer hired a team of legal engineers from Norm Ai to build agentic workflows under its chief AI officer. Business Wire via Yahoo Finance · Global Legal Post

Quick Hits

The wider field, one line each. Funding figures and product claims are the companies' own.

Agent security funding

01Bay Security raised $50 million, including a $37 million Series A led by Battery Ventures, to stop rogue agent actions. The Next Web
02WhiteMatter raised a $20 million Series A led by Blumberg Capital for AI agent security. SecurityWeek
03Metano came out of stealth with a $12 million seed led by Sierra Ventures to secure enterprise agents. Metano
04Disseqt AI raised $8 million led by Gresham House Ventures to test and police enterprise agents. Tech Funding News
05Mitiga launched AI Agent Security, which finds agents across cloud and SaaS and can contain one by revoking its tokens. Mitiga

Agents at work

06Payhawk raised a $280 million Series C at a $2 billion valuation, led by Goldman Sachs Alternatives, to put agents into finance work. GlobeNewswire via Yahoo Finance
07Linq raised $90 million led by Andreessen Horowitz for infrastructure that lets AI agents text, call and use computers. PR Newswire
08Pulley raised $64 million across its Series A and B for AI permitting in commercial real estate. Business Wire
09LATO raised about $4.7 million, co-led by Air Street Capital, for agents that run expert interviews in private-equity due diligence. The Recursive
10Modern Industrials raised a $6 million seed co-led by Glasswing Ventures and TMV for custom agents at industrial manufacturers and distributors. GlobeNewswire

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

$177B

Enterprise AI spending by September 2026, about 5.4 times the $33 billion of 2025, by Menlo Ventures' estimate.

Anthropic, OpenAI and Google took about 69% of deployment spending. Menlo built the figure bottom-up and invests in some of the companies it counts, so treat it as a direction of travel.

Source: HPCwire

Counter-Signal

Governance

A kill switch doesn't decide who pays

This week's deals suggest agent risk is being handled: buy a guard and a kill switch, then move on. The contracts written the same week point the other way. Citi and PNC now put their customers on the hook for what their agents do. AWS lets agents buy software and leaves it to each buyer's policy which purchases a person checks.

A security platform can stop an agent mid-action. Losses that a bank or vendor has already written onto your side of the contract sit with legal and procurement, and that work is moving faster than the tooling. This quarter, the agent clauses in terms you've already accepted may matter more than the platform you buy.

Sources: TechCrunch · Fortune

From the Field

When we review agent programs, ownership is the column most often left blank.

Teams can usually name the developer who built an agent and the platform it runs on. The budget its actions draw on, and the person who answers when its access goes wrong, are more often a guess. That gap was easy to live with while agents mostly read and summarized. It gets expensive once they start acting with money and credentials.

The security deals will help with the watching. Palo Alto and CrowdStrike can see what an agent does far better than most companies can on their own. Watching is the easier half of the job. Deciding who answers for an agent is a management call, and no acquisition this week comes with one.

In the programs that run well, each agent has one named owner who holds its budget and can switch it off.

Most of the agents we see got a budget before they got an owner.

Let's create value with AI,
AK

The Agentic Enterprise

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

Anthropic, OpenAI, Google, Microsoft/Azure, AWS, NVIDIA, Databricks and Snowflake are Spearhead technology partners; this edition covers several of them, including AWS’s agent purchasing, Microsoft’s Transparency Initiative, OpenAI’s ChatGPT for Sales and Anthropic’s IPO close, and applies the same scrutiny to each. The Big Story concerns Palo Alto Networks, not a listed partner. This edition was produced using Anthropic’s Claude with editorial direction from AK. Sources are linked inline.

© 2026 Spearhead. All rights reserved.