Google Cloud shipped IAM Helper, an agent-native layer over Google Cloud IAM that lets coding agents diagnose 403s and request just-in-time privileged access with policy simulation before any change lands
AspireAI community · Google Cloud Blog
Google Cloud introduced IAM Helper on October 27, 2026 — a suite of agent-facing building blocks that makes identity and access management callable by autonomous code. The suite packages end-to-end IAM workflows as Agent Skills, an open standard, so agents can resolve permissions without an admin in the loop. Diagnostics and policy simulations run automatically; any change to access policy pauses for human-in-the-loop confirmation based on customizable risk tiers. The shipped skills include troubleshooting (identify the exact missing permission behind a 403 PERMISSION_DENIED), privileged access management (request a one-hour JIT grant), policy management (generate and validate IAM v1 allow bindings and IAM v2 deny rules, blocking public access and broad basic roles like roles/owner and roles/editor before they are applied), and a policy simulator that tests proposed changes against 90 days of historical access logs before rollout. Skills work across Gemini CLI, Gemini Cloud Assist, ADK, Claude Code, and Cursor. The honest limitation: this is an orchestration layer over existing Google Cloud IAM engines — Policy Intelligence and Privileged Access Manager — not a new identity model, and the human-in-the-loop gate means the autonomous path stops at every policy write.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.