qa
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes multiple vendor-provided binaries (e.g.,
gstack-skill-start,gstack-question-log,gstack-paths) located in~/.claude/skills/gstack/bin/for initialization, state management, and logging. - [EXTERNAL_DOWNLOADS]: The skill detects project runtimes and offers to install external test frameworks (e.g., vitest, pytest, rspec) via standard package managers. This requires explicit user approval.
- [INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection. Ingestion points: Processes untrusted content from web pages (HTML, text, screenshots) and external search results during orientation and exploration phases. Boundary markers: Present (
═══ BEGIN/END UNTRUSTED WEB CONTENT ═══). Capability inventory: PossessesWrite,Edit, andBashcapabilities allowing for repository modification as part of the fix loop. Sanitization: Instructions state that all page-returned data is untrusted content, though explicit escaping of all interpolated data is not detailed. - [DYNAMIC_EXECUTION]: The skill processes dynamic instruction blocks (
GSTACK_INSTRUCTION_BEGIN) emitted by its startup tool. It includes verification logic to ensure these instructions only originate from a direct, local tool call with a matching session ID. - [DATA_EXFILTRATION]: Local telemetry and analytical insights (Eureka logs) are stored in JSONL format within the vendor's state directory (
$GSTACK_STATE_ROOT/analytics/).
Audit Metadata