structured-output-fabric

Pass

Audited by Gen Agent Trust Hub on Mar 12, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it concatenates untrusted input text directly into LLM prompts.
  • Ingestion points: Untrusted source text is ingested and processed in SKILL.md (Step 4 and 5) and examples/example_workflow.sh (Step 3).
  • Boundary markers: The skill uses simple newline separators and plaintext headers (e.g., '## User Messages:'), which do not provide robust isolation between instructions and data.
  • Capability inventory: The skill utilizes shell command execution (bash, jq, fabric, sed) and filesystem writes, as documented in SKILL.md, examples/example_workflow.sh, and scripts/extract_json_from_llm.sh.
  • Sanitization: No input sanitization or escaping is performed before the data is included in the model prompts.
  • [COMMAND_EXECUTION]: The skill performs automated execution of system commands and local scripts to achieve its workflow.
  • It invokes the fabric CLI tool for both prompt improvement and querying the LLM.
  • It executes the local bash script scripts/extract_json_from_llm.sh to handle structured data extraction.
  • It relies on standard utilities including jq for output validation and sed for text manipulation within its automated steps.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 12, 2026, 12:05 PM