The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill enables the agent to read data from external sources like Cloud Storage (cat/cp), Secret Manager (access latest), and Cloud Logging (read/tail). If an attacker can write to these cloud resources, they could potentially embed malicious instructions for the agent.
Ingestion points: storage.md (gcloud storage cat), iam.md (gcloud secrets versions access), compute.md (gcloud logging read), and data.md (bq head).
Boundary markers: Absent; the skill relies on standard command output without specific delimiters to isolate untrusted data.
Capability inventory: High; the skill includes commands for resource deletion, IAM policy modification, firewall configuration, and service deployment.
Sanitization: Absent; there are no instructions for the agent to validate or sanitize content retrieved from the cloud environment.
[Unverifiable Dependencies & Remote Code Execution] (LOW): The skill suggests installation via the scoop package manager and references Google-maintained GitHub Actions.
Evidence: scoop install gcloud in SKILL.md and google-github-actions/auth@v2 in automation.md.
[TRUST-SCOPE-RULE]: Google and its associated repositories are trusted organizations, and scoop is a standard tool for the platform, downgrading the download finding to LOW/INFO.

gcloud