aks-gpu-inference
Pass
Audited by Gen Agent Trust Hub on Oct 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- Indirect Prompt Injection Surface: The skill instructs the agent to retrieve and analyze container logs which are considered untrusted external data.
- Ingestion Points: In
references/gpu-observability.md, the commandkubectl logs -n <namespace> <pod> -c <container> --previous --timestampsis used to collect model and runtime logs. - Boundary Markers: There are no explicit delimiters or instructions provided to the agent to ignore potentially malicious content within these logs.
- Capability Inventory: The skill has access to
kubectlandazCLI tools for cluster and cloud resource inspection. - Sanitization: No specific sanitization or filtering of the log output is mentioned before the agent processes the content.
- Command Execution Capabilities: The skill utilizes shell-based tools to gather evidence from the cluster environment.
- Execution Method: Commands such as
kubectl get,kubectl describe, andaz aks showare defined across all reference files to inspect pod status, node configurations, and workspace states. - Context: While these commands are essential for the skill's diagnostic purpose, they interact with sensitive infrastructure. The skill mitigates risk by including a 'Host Capability Gate' in
SKILL.mdthat requires the host to authorize these specific commands and by emphasizing a read-only approach for all evidence collection.
Audit Metadata