tilegym-converting-cutile-to-julia
Warn
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The workflow documentation in
translations/workflow.mdexplicitly directs the AI agent to bypass standard confirmation protocols using instructions like 'Auto-proceed: Move to next phase automatically after success — NO user confirmation needed' and 'DO NOT ask "should I proceed?" — execute the full workflow end-to-end.' This overrides the agent's safety protocols regarding human-in-the-loop validation. - [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface because it ingests untrusted user-provided Python code and performs code generation and shell-based testing without explicit boundary markers or input sanitization.
- [COMMAND_EXECUTION]: Documentation in
references/debugging.mdincludes the commandrm -rf ~/.julia/compiled/cuTile*for clearing compilation caches; use of recursive deletion with wildcards is a destructive pattern that may lead to unintended data loss. - [EXTERNAL_DOWNLOADS]: The environment setup requires downloading dependencies from the official Julia package registry using
Pkg.instantiate(), which is a standard but external network operation.
Audit Metadata