Photo: Tima Miroshnichenko / Pexels
The beta product isolates agent-generated code using WebAssembly, arriving alongside a new managed inference service for open-weight models
AI agents are getting good at writing their own code. The harder question is where that code is allowed to run.
Akamai’s answer, unveiled on Tuesday, is a sealed box. The company launched Agent Sandbox, an isolated environment where code produced by an AI agent can execute without reaching the rest of a company’s systems.
The product is in beta. It shipped alongside Akamai Inference, a managed service for running open AI models on Akamai’s cloud, which is also in beta.
The timing is not subtle. Palo Alto Networks, Cisco, CrowdStrike, Okta and Zscaler have all either acquired or rolled out AI-agent security products this year, and Akamai is now stepping onto a crowded stage.
How the sandbox works #
Technically, the isolation is built on WebAssembly. Akamai uses the open-source Spin framework, running inside its Akamai Functions product, to wall off each piece of agent-generated code.
WebAssembly is a compact format for running programs in a tightly restricted environment. Put simply, the code gets only the permissions it is explicitly handed, and nothing more.
Akamai says the sandbox places strict limits on outbound access. It also caps how much compute and memory a given task can consume.
Those two controls matter for different reasons. Outbound limits stop a misbehaving agent from quietly sending data somewhere it shouldn’t. Resource caps stop a runaway script from eating the server it lives on.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The inference side of the launch #
Akamai Inference is the other half of Tuesday’s announcement. It is a managed service for open-weight models, meaning AI models whose underlying parameters are publicly released rather than locked behind a single vendor.
Customers reach those models through APIs that are compatible with OpenAI’s.
Organizations can deploy models on a serverless setup or on dedicated GPU resources.
At launch, the service supports twelve open-weight models. They range from 8 billion parameters at the small end to 2.8 trillion at the top.
The lineup includes NVIDIA’s Nemotron 3 Ultra, OpenAI’s gpt-oss-120b and Google’s Gemma 4.
Background: Akamai’s AI push #
Akamai is best known for content delivery and security, but it now positions itself as an AI-focused cloud provider. Both new products sit inside its AI Cloud platform.
The company has reported securing over $14 billion in multi-year cloud contracts.
This is also not Akamai’s first move on agent security. In June 2026, it introduced an agentic security framework in partnership with Visa and other companies.
Tuesday’s launch extends that work from policy into infrastructure. A framework describes how agents should behave, while a sandbox physically limits what they can do.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our