skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 6eedbdc2f44c
- Files read
- 40
This project provides seven Google ADK skills that guide AI coding agents through designing, scaffolding, coding, evaluating, deploying, publishing, and monitoring agents on Google Cloud using agents-cli. The biggest risks are an observability guide that instructs granting over-broad Cloud Storage and BigQuery admin permissions to a deployed agent's service account, and recipe commands that reuse predictable folders in /tmp on shared computers without checking ownership.
What it can reach
- Shell commands
- Instructs the coding agent to run agents-cli, uv, go, git, terraform, gcloud, bq, gsutil, kubectl, and make commands to scaffold, test, evaluate, provision infrastructure, and deploy agents.
- Files
- Reads and writes project source code, .agents-cli-spec.md, .env files, evaluation datasets and traces, Terraform configurations, and temporary clones or reference scaffolds under /tmp.
- Network
- Fetches ADK documentation from adk.dev, clones recipe and template repositories from GitHub, installs Python and Go dependencies, and calls Google Cloud and Gemini platform APIs.
- Secrets
- Uses Google Cloud credentials, GitHub CLI authentication, GCP Secret Manager secrets, and project .env variables that are propagated during deployment.
Findings
-
MediumPredictable /tmp paths reused without checking ownership
skills/google-agents-cli-adk-code/references/recipes.md:16
When studying reference recipes, the skill instructs the coding agent to check whether /tmp/adk-recipes already exists and, if so, skip cloning and immediately read AGENTS.md and code from that directory (and google-agents-cli-scaffold similarly creates temporary projects at /tmp/ref-project, which switches into enhance mode if the folder already exists). On a shared multi-user machine where /tmp is world-writable, another local user can pre-create /tmp/adk-recipes or /tmp/ref-project with tampered agent instructions or Makefiles that your coding agent will read and copy into your project.
Fix: Create unique temporary directories with mktemp -d instead of using fixed paths like /tmp/adk-recipes and /tmp/ref-project, or verify directory ownership and Git remote origin before reading files.
-
LowObservability guide instructs granting over-broad Storage Admin and BigQuery Data Owner roles to runtime agent
skills/google-agents-cli-observability/SKILL.md:36
When adding observability to an agent that was already deployed imperatively with agents-cli deploy, the skill tells the agent or user to grant roles/storage.admin (and roles/bigquery.dataOwner) to the deployed agent's runtime service account alongside project-level roles without scoping them to a specific bucket or dataset. Writing completions to the logs bucket only requires roles/storage.objectCreator on that bucket, as the troubleshooting table later notes. Granting roles/storage.admin, especially at the project level, allows a compromised or prompt-injected production agent to read, modify, or delete objects across Cloud Storage buckets—including compliance logs and Terraform state—and change bucket IAM policies.
Fix: Replace roles/storage.admin with roles/storage.objectCreator bound only to the specific logs bucket, and scope BigQuery permissions to the single telemetry dataset instead of granting project-wide admin or owner roles.
-
LowUnpinned recipe clones and uncurated community contrib recipes
skills/google-agents-cli-adk-code/references/recipes.md:95
The recipe guide clones google/adk-recipes from its mutable default branch without pinning a release tag or commit SHA (including cloning immediately without user approval on a 'how' question), and directs the coding agent to check contrib/—which is community-contributed and not curated by the agents-cli team—when no core recipe matches. Because the scaffolding skill tells the agent to read a recipe's instructions, copy its code and Terraform, and run its Makefile targets, a compromised or malicious community contribution in contrib/ could inject instructions into the coding agent or run unintended commands.
Fix: Pin recipe clones to a specific release tag or commit SHA, ask for user confirmation before cloning external repositories or executing recipe Makefile targets, and explicitly warn the agent to treat contrib/ code and markdown as untrusted.
-
LowDeploy skill mandates --no-confirm-project without verifying the resolved GCP project with the user
skills/google-agents-cli-deploy/SKILL.md:149
When --project is omitted, agents-cli deploy resolves the Google Cloud project automatically from ambient settings and normally prompts for confirmation so you do not deploy to the wrong project by mistake. This skill instructs the coding agent that it MUST pass --no-confirm-project when relying on automatic resolution, while its human approval step ('Ready to deploy to dev?') does not show which GCP project was resolved. If local gcloud settings or .env point to an unintended or production project, the agent will deploy or modify environment variables and secrets there without confirming the target project ID.
Fix: Instruct the coding agent to inspect and show the target GCP project ID in the human approval prompt, then pass --project <PROJECT_ID> explicitly rather than bypassing confirmation with --no-confirm-project.
