News

Tools

CONFIRMED: Google Gen AI SDK 2.29.0 simplifies blocking agent network access

The October 7 update lets developers block outbound networking inside the environment configuration. It provides a more explicit control for code-running agents, not a new Gemini model or a complete security guarantee.

Ilustração editorial de blocos de código coloridos conectados por portas de rede e uma fechadura, representando o controle de acesso de agentes.

CONFIRMED

Google released version 2.29.0 of the Google Gen AI SDK, its software development kit for Python, on October 7, 2026. The central change lets an environment’s network configuration explicitly declare that no outbound connections are authorized. For businesses using agents that can execute code, this makes it easier to separate tasks that require internet access from tasks that should operate only on supplied files.

One word that closes the gate

The update accepts disabled in the allowlist field inside the network object. An allowlist specifies destinations approved for network connections. Here, disabled does not open all destinations or turn off a protection: according to the official commit, it blocks all terminal network egress. The previous form, setting network directly to disabled, continues to work. The benefit is an additional way to express policy inside the configuration, not the introduction of network blocking to the product.

A concrete business use case

Consider a small Brazilian company giving an agent a sales spreadsheet to produce charts and check totals. If the task can be completed using the supplied data, the terminal has no reason to browse websites or send information to other services. Explicit blocking helps constrain that pathway. An automation that retrieves inventory from an external system needs a different policy. The decision should follow the task, rather than making broad access the default simply because it is convenient.

A control is not a complete architecture

The feature concerns outbound networking from the configured environment. It does not demonstrate that the entire application is offline, that data no longer reaches the model provider, or that other external tools have been blocked. It also does not replace file permissions, code review, personal-data handling, or execution limits. Before upgrading a critical integration, a business should test behavior on the service and account it actually uses. Changes to types and class paths in the same commit deserve attention in projects that import those components directly.

Examples are not model launches

The release also adjusts a warning about automatic function calling, the mechanism through which the client executes available functions, so it appears only when functions are configured. It updates examples using Gemini identifiers as well. Changing examples does not prove a new model launch or universal access. Version 2.29.0 is listed on the Python Package Index, the PyPI package catalog, confirming distribution beyond the repository tag. This reporting did not conduct an environment security test or measure operational savings.

A modest but useful evolution

This canonical article was updated on October 8. Version 2.28.0, released on October 2, had added identifiers and a continuation token to Interactions, the application programming interface for stateful conversations and tools. Version 2.29.0 brings a different material change: clearer network policy. Our editorial view is that this control matters more to automation reliability than a new name in a model list. Its benefit depends on properly configuring, testing, and restricting the task, rather than simply installing the package.

Sources

Google: Python SDK 2.29.0 release; Google: network egress blocking change; PyPI: google-genai 2.29.0 distribution; Google: version 2.28.0 history.