Akamai launches Agent Sandbox to limit what AI code can do

Choosing where an AI model runs does not settle what its generated code can access. Akamai’s new services address both decisions, with preview limits that matter for developers.

Akamai announced Inference and Agent Sandbox on October 27, separating two tasks often bundled together in discussions about AI deployment. Inference serves models. Agent Sandbox provides a restricted environment for agents and the code they generate. Both are being introduced in beta, according to the company. Akamai Technologies Inc.

Consider a hypothetical sales assistant. A user asks it to compare this month’s revenue with last month’s. The model writes a short Python calculation. The application then runs that code and passes the result back to the assistant for an explanation.

There are separate decisions along that route. The company chooses which model receives the request, which data goes into the calculation, and which systems the generated code can access. A script that only needs two sales totals does not need credentials for the entire customer database.

These permissions can differ even when one cloud provider supplies both the model service and the execution environment. Hosting a model in a chosen region does not, by itself, define what a generated program may do.

According to Akamai’s Agent Sandbox product page, the agent runs in a Spin-based WebAssembly sandbox on Akamai Functions. Generated JavaScript or Python runs inside an additional restricted runtime. Outbound network access is blocked by default, and execution limits are enforced by the runtime.

The agent can connect to models through an OpenAI-compatible API, including external providers. Using the sandbox therefore does not require hosting the model with Akamai.

The underlying idea is explicit permission. Spin’s application documentation describes how developers grant access to files, storage and network destinations. Its HTTP documentation says requests to hosts outside the permitted list fail. These documents explain Spin’s general controls; they do not establish that every configuration option is exposed identically in Akamai’s managed service. Spin Docs

Akamai’s isolation and containment statements remain vendor claims. AI World Today has not independently security-tested this preview.

The product FAQ calls Agent Sandbox a tech preview for intermittent agents: each step completes in about a minute, then the agent stops until its next instruction. Support for agents that remain alive with state for days or weeks is described as future functionality.

The page specifies deployment across Akamai Cloud core regions. Akamai’s wider global edge footprint should not be read as a list of locations where this sandbox is available.

Akamai Inference provides managed model access through an OpenAI-compatible API. Akamai handles the underlying serving infrastructure. Its product page describes two deployment options:

The pricing FAQ says there is no charge during the inference beta and that billing begins at limited availability. That is a temporary beta condition, not a promise of permanently free inference. It should not be extended to every associated cloud resource.

As of October 27, the product page does not give numerical token or GPU-hour rates, a billing-start date, or a general-availability date. Those details still need confirming before a production cost comparison.

The practical implication is that isolation and correctness require separate checks. A restricted script can still calculate the wrong figure, process the wrong records or produce a misleading result. Limiting where it runs does not validate its answer.

Data handling also spans more than the execution environment. In the sales example, the application may send the calculation’s output back to a remote model for interpretation. Blocking direct network calls from generated code would not remove that deliberate data path. Developers still need to decide what information the model receives.

AI World Today’s Llama API guide explains the choice between a hosted model, a specialist inference provider and a self-hosted API. Akamai’s announcement adds a separate question to that decision: once the model writes executable code, which environment runs it, and what permissions does that environment grant?