Palo Alto Networks will acquire Applied Compute for $4.5 billion, extending its reach into the infrastructure that runs AI Agents. The deal signals a broader shift in how enterprise security must account for dynamic, tool-using AI workloads.
Palo Alto Networks has announced a $4.5 billion acquisition of Applied Compute.
The company says the deal will combine its security capabilities with open cloud infrastructure.
This is more than a routine product-line acquisition.
Palo Alto Networks has built its business around firewalls, zero-trust access, and policy management. By adding Applied Compute's cloud infrastructure capabilities, it is targeting the compute, networking, and runtime environments used by AI workloads.
Its ambition extends beyond controlling “which users can access which applications.” It also wants to govern “where AI Agents run, what they can call, and which data they can see.”
Why This Deal Matters
Traditional cybersecurity products are usually organized around boundaries.
A company has an office network, data centers, and cloud accounts. Firewalls inspect traffic, EDR monitors endpoints, and IAM manages identities. Once those products are deployed, the enterprise continues buying compute from AWS, Azure, or Google Cloud.
This division of labor has an obvious weakness: security vendors control policy, but not the execution environment.
AI Agents amplify that problem. An Agent does more than answer questions. It can call APIs, read databases, update ticketing systems, and even perform write operations based on business rules.
Temporary files, containers, network connections, model caches, and tool credentials created by an Agent runtime can all become new attack surfaces.
Applied Compute provides more than a bundled offering. It gives Palo Alto Networks cloud infrastructure capabilities. The company wants to place security policies directly into the infrastructure layer so identity, networking, and workload execution can operate through a unified control plane.
That is different from acquiring isolated security tools. In the past, a security company might buy a detection capability and add it to its existing product portfolio. This time, Palo Alto Networks is acquiring the foundation on which Agents run.
Why AI Agents Need a New Security Boundary
AI workloads do not behave exactly like traditional enterprise applications.
An Agent task might begin by asking an LLM to generate a plan. The Agent then uses tools to access internal services, creates a container or virtual machine, runs code, stores intermediate state, and writes the result back to a business system.
Any step in that sequence can fail or be abused.
For example, an attacker could inject content into an Agent's prompt. Tool credentials may be overprivileged. Two tasks may share the same runtime environment. Test code may gain access to production keys.
Historically, these problems have fallen across IAM, container security, cloud security, and developer tooling teams.
The core change brought by open cloud infrastructure is that the security boundary now extends from “networks and accounts” to the “execution environment.”
Applied Compute gives Palo Alto Networks more direct control over:
- Which clusters run Agent workloads;
- How containers, virtual machines, and images may be created;
- Network access between workloads;
- Runtime identities and policies within a single platform;
- Checkpoints between task execution and tool calls.
These capabilities will not emerge automatically when the transaction closes. Applied Compute was built for open cloud infrastructure, while Palo Alto Networks has traditionally focused on policy and protection.
How the two product sets, teams, and commercial agreements will come together remains unclear. Palo Alto Networks' future disclosures will be important.
Official press release:
Palo Alto Networks to Acquire Applied Compute, Bringing Secure, Open Cloud Infrastructure to the AI Era
Agent Security Requires Four Layers
Agent security cannot be solved at the application layer alone.
Consider a customer-service Agent that can look up orders, change addresses, and issue refunds. On the surface, it is an LLM with a set of function-calling tools.
The real risk boundary depends on four layers:
- Identity: Which identity does the Agent use to access systems?
- Authorization: Which resources can that identity read or modify?
- Execution: Where do code and tools run?
- Audit: After an unusual action, can investigators reconstruct the full chain of events?
Many enterprises already own IAM, SIEM, and cloud security products, yet still lack controls at the execution layer. When an Agent dynamically creates workflows from prompts, administrators may not even know what it ran.
This also creates new challenges for the MCP and tool ecosystem. An MCP Server can expose files, databases, and business systems as tools that Agents can call.
As the number of connections grows, managing credentials, authorization scope, and runtime isolation becomes increasingly difficult without automation.
Palo Alto Networks' vision is to place Agent identity, permissions, networking, and runtime environments on one platform. That would spare enterprises the need to assemble separate IAM, container security, and cloud firewall systems.
This does not mean existing tools will disappear. Agent frameworks still need sandboxes, permission declarations, and observability. A major security vendor can provide a more complete infrastructure view without replacing the code developers use to build Agents.
Three Challenges After the Acquisition
Conflict of Interest
Palo Alto Networks sells a security platform, while Applied Compute provides open cloud infrastructure. Enterprise customers may want to retain AWS, Azure, Google Cloud, or their own Kubernetes deployments while purchasing security products from the combined vendor.
If the integration goes too far, customers may worry that gaining a more complete security platform will require them to migrate workloads or accept a more closed ecosystem.
Similar concerns have followed the cloud strategies of VMware, Cisco, and other large technology vendors. A stronger control plane does not eliminate customer concerns about choice, compatibility, and total cost of ownership.
Pricing and Migration
Will Applied Compute continue to support multicloud and customer-managed environments? Will the security capabilities be priced independently rather than bundled with an entire cloud resource package?
And can existing APIs, IAM systems, and policy models be adapted to Agent workloads?
Developer Habits
Enterprise security leaders may prefer a unified console, but Agent engineers typically begin with tools such as LangGraph, the OpenAI SDK, or the Anthropic API.
If the new platform cannot work with existing code, tool protocols, and CI/CD pipelines, it is unlikely to become the default runtime environment.
The $4.5 billion purchase therefore buys a strategic position, not a guaranteed product outcome. The next disclosures to watch concern product scope, pricing, customer migration, and support for an open ecosystem.
Editorial Perspective
The significance of this deal lies not in a security vendor adding another cloud product, but in the way AI Agents are pushing security responsibility down into the execution layer. Enterprises used to buy separate identity, networking, and monitoring tools for each system. Once Agents dynamically create tasks, call tools, and modify data, those boundaries must operate together.
Palo Alto Networks is trying to merge policy with the compute foundation in a bid to control the Agent runtime. It is unlikely to displace AWS or Azure on its own; a more plausible outcome is a cross-cloud security and execution control layer. Even so, this acquisition differs from earlier purchases of isolated tools. The company previously bought inspection capabilities; it now wants to control the workloads themselves. Integration timelines, customer migration costs, and whether “open infrastructure” can withstand commercial pressure remain uncertain.