ccc-testing
Warn
Audited by Gen Agent Trust Hub on Apr 1, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The script
webapp-testing/scripts/with_server.pyexecutes server commands usingsubprocess.Popenwithshell=True. This pattern is risky if user-supplied input is passed to the server arguments without strict sanitization. - [EXTERNAL_DOWNLOADS]: The
plankton-code-qualityskill downloads external code from a community GitHub repository (alexfazio/plankton) usinggit clone. While Plankton is a known community tool, it is not an officially trusted vendor repository. - [EXTERNAL_DOWNLOADS]: The
qaandqa-onlyskills include instructions to download and execute the Bun installer frombun.shusing thecurl | bashpattern. This is a well-known service, but the execution method bypasses standard package manager safety checks. - [DATA_EXFILTRATION]: The
Preamblesections inqa/SKILL.mdandqa-only/SKILL.mdcontain shell commands that automatically collect and log telemetry. This includes the name of the current repository (basename "$(git rev-parse --show-toplevel)"), session IDs, and skill usage data, which are stored in~/.gstack/analytics/and sent to an external telemetry service. - [DATA_EXFILTRATION]: The skill implements session tracking by creating files in the user's home directory (
~/.gstack/sessions/) named after the parent process ID ($PPID). - [PROMPT_INJECTION]: The
qaandqa-onlyskills utilize a 'Completeness Principle' that strongly instructs the agent to always prefer complete implementations over shortcuts, even when not explicitly requested by the user, which could lead to unexpected resource consumption or code changes.
Audit Metadata