The Agent Skills Directory

[Indirect Prompt Injection Surface]: This skill ingests untrusted data from service logs and pod logs (e.g., via kubectl logs and az webapp log tail) as seen in files such as scripts/pod-evidence.sh and references/app-service/README.md. This data enters the agent context without explicit boundary markers or sanitization. The capability inventory includes the Azure CLI, kubectl, and various diagnostic scripts used to process this data.
[Privileged Command Execution]: The run-ig.sh and run-ig.ps1 scripts facilitate the execution of kubectl debug --profile=sysadmin for node-level observability. This represents a privilege escalation surface as it provides root-level access to cluster nodes. The documentation includes explicit instructions to seek user approval before performing these actions.
[Instruction Override]: The SKILL.md file contains emphatic directives such as "AUTHORITATIVE GUIDANCE — MANDATORY COMPLIANCE". While intended to define a systematic diagnostic workflow, this pattern of overriding instructions is a common indicator of prompt injection techniques.
[Access to Sensitive Data]: Troubleshooting guides, such as the one for Function Apps in references/functions/README.md, include commands for retrieving application settings and connection strings. While necessary for debugging, this capability provides a pathway for accessing potentially sensitive environment variables.

azure-diagnostics