luke-mode
Warn
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_CONTEXT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill's setup and sync playbooks automate the installation of external AI agent skills from repositories such as 'mattpocock/skills' and 'backnotprop/pstack' using the 'skills' CLI (npx skills@latest).
- [COMMAND_EXECUTION]: Multiple Python scripts included in the skill ('device_claim.py', 'publish_evidence.py', 'role_agents.py') execute shell commands using the subprocess module to interact with system tools like 'git', 'ps', and mobile device managers ('xcrun simctl', 'adb').
- [DYNAMIC_CONTEXT_INJECTION]: The skill utilizes a Claude Code-specific 'PreToolUse' hook to execute 'scripts/role_agents.py hook' before every invocation of the Agent tool. This mechanism is used to intercept and potentially deny tool calls based on configured model routing policies.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including ticket descriptions, specifications, and PR bodies, which are subsequently interpolated into subagent briefs without explicit sanitization or boundary enforcement logic in the provided instructions.
- [DATA_EXFILTRATION]: The 'publish_evidence.py' script is designed to push local files (such as screenshots or recordings) to a remote Git repository under the 'refs/evidence/' namespace.
Audit Metadata