Google Cloud launches IAM Helper to let agents manage access

Google Cloud launched IAM Helper on 11 November 2026, a suite of agent-facing tools that lets autonomous software agents diagnose and manage identity and access tasks across Google Cloud's existing IAM stack, including Policy Intelligence and Privileged Access Manager.

The product targets a specific and growing operational problem. Businesses increasingly deploy autonomous agents to provision cloud resources, write infrastructure code and deploy services. Those agents routinely hit permission barriers inside continuous integration and deployment pipelines. IAM Helper positions itself as the orchestration layer that automates diagnosis of those failures and routes higher-risk policy changes to humans for approval.

The stakes for the market are direct. Agent-driven development has pushed cloud teams into a trade-off between velocity and least-privilege discipline. Google Cloud's answer is architectural: keep simulations and diagnostics automatic, pause policy changes for human-in-the-loop confirmation, and handle heavy policy evaluation on dedicated backend endpoints rather than inside large language model contexts.

What does IAM Helper actually do?

The launch includes several discrete functions.

One function covers policy management. It generates and validates IAM v1 allow bindings and IAM v2 deny rules in formats that include JSON, YAML and Terraform HCL, and blocks risky configurations such as public access settings or broad basic roles before they reach production.

A second function addresses permission errors directly. It identifies the missing permissions or blocking rules behind 403 errors. The troubleshooting workflow serves two audiences: developers resolving issues themselves, and administrators tracing inherited allow and deny rules across organisational structures.

A policy simulator lets engineers assess the effect of proposed changes before applying them. The simulator tests changes against 90 days of historical access logs to determine whether workloads or services would be disrupted.

A just-in-time privileged access workflow, tied to Privileged Access Manager, supports temporary production access for responders or maintenance agents. Approvals, audit records and automatic revocation are built into that process.

Google Cloud set out two concrete usage examples. In the first, a developer's coding agent hits a 403 error during an application deployment, identifies a missing Pub/Sub permission, and requests a one-hour just-in-time access grant instead of creating a support ticket. In the second, an infrastructure agent that meets a permission block during a routine deployment pipeline generates a revised least-privilege policy, checks it against historical logs, and routes the change to an administrator for approval rather than letting the build fail outright.

The second example is the product's thesis in miniature: the agent does not widen permissions to unblock itself. It proposes a narrower policy, verifies it against evidence, and defers the decision.

Where does the tool run?

IAM Helper ships packaged on the open Agent Skills standard and works across several coding and assistant environments. Google Cloud lists Gemini CLI, Gemini Cloud Assist, ADK, Claude Code and Cursor as supported surfaces.

That cross-environment reach matters. The tool is not confined to Google's own assistant products; it extends to Anthropic's Claude Code and Anysphere's Cursor, the coding environments where much agentic development already happens.

How does it reduce cost and delay?

IAM Helper changes where policy evaluation work happens. Google Cloud states that the product is intended to reduce the amount of policy evaluation performed by human administrators and by large language models. Instead, IAM Helper sends orchestration tasks to Google Cloud's backend IAM services.

That design should reduce token usage and operational delays. It is a concrete comparison worth holding onto: an LLM reading role hierarchies and group memberships burns context on data that a dedicated endpoint returns in a structured call. Google Cloud moved that work off the model and onto the endpoints built for it.

What sits underneath: a Google-managed MCP layer

Below the higher-level skills sits a set of Google-managed MCP tools connected to core IAM endpoints: Policy Troubleshooter, Policy Analyser, Policy Simulator and IAM APIs. Google exposes these as read-only atomic functions that agents can discover and invoke directly when developers need more customised workflows.

The practical consequence is that teams do not need to build their own wrappers for individual APIs when they want agents to check permissions, list roles or traverse resource hierarchies. The design also lets customers combine Google Cloud's IAM functions with internal ticketing systems, compliance checks and deployment pipelines.

Heavy evaluation tasks — hierarchy traversal, group expansion, rule analysis — run on dedicated IAM endpoints spanning IAM v1 allow roles, IAM v2 deny, Policy Troubleshooter and Policy Analyser.

Why the read-only and write separation matters

The suite's core safety property is the line it draws between reading and writing. Simulations run automatically. Access policy updates pause for human confirmation, with the threshold governed by customisable risk tiers.

That split addresses what Google Cloud frames as a central challenge in agent-driven development: removing access bottlenecks without widening permissions. An agent that can diagnose its own 403 error is useful. An agent that can silently rewrite IAM policy is a liability. IAM Helper gives the first capability freely and gates the second.

The launch reflects a broader shift in cloud operations. Companies want to grant software agents more autonomy while preserving least-privilege controls, and the identity layer is where that tension is sharpest. By placing simulations ahead of changes and requiring approval for higher-risk updates, Google Cloud has staked out a position other cloud providers will likely have to match as agents take over more of the deployment pipeline.