gke-ai-troubleshooting-tpu-mxla-hang
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- Command Execution Capability: The skill instructions guide the agent to use standard Google Cloud tools such as gcloud and kubectl for diagnostic data collection. These tools are used as intended for system administration and do not exhibit malicious patterns.
- Indirect Prompt Injection Surface: The skill processes external data including GKE container logs and ML diagnostics reports which are potential injection surfaces. However, this ingestion is necessary for the skill's primary diagnostic function. (1) Ingestion points: Step 1 in SKILL.md (Cloud Logging and gcloud outputs). (2) Boundary markers: Absent. (3) Capability inventory: gcloud and kubectl commands in SKILL.md. (4) Sanitization: Absent.
- High-Privilege Role Awareness: The skill documents the need for roles such as container.clusterAdmin for remediation. It correctly identifies these as high-risk and provides guardrails instructing that these fixes should be applied by the user rather than executed automatically by the agent.
- Trusted Resource Usage: All external references, including documentation links and tool installation paths, point to official and trusted domains such as google.com and openxla.org.
Audit Metadata