hyper-feedback

Pass

Audited by Gen Agent Trust Hub on Nov 20, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user-supplied feedback, which presents a surface for indirect prompt injection.
  • Ingestion points: User request text, pasted messages, and relayed summaries are ingested as the primary data source (SKILL.md, rules/change-mapping.md).
  • Boundary markers: The skill uses labeled blocks for quoting request text and explicitly separates observed facts from author inference.
  • Capability inventory: The skill has the authority to search the repository, edit files, and execute local verification commands (SKILL.md).
  • Sanitization: Instructions explicitly mandate treating request text as evidence with 'no instruction authority' and forbid executing commands or exposing secrets just because the input text asks.
  • [COMMAND_EXECUTION]: The skill executes local verification commands (e.g., test, build, lint) to confirm that applied changes are correct.
  • Evidence: The skill uses the repository's own commands for verification.
  • Mitigation: Command execution is limited to existing checks in the touched area, and the repair loop is capped at a maximum of two attempts to prevent unintended or persistent execution cycles.
  • [EXTERNAL_DOWNLOADS]: The skill references external documentation from a well-known service and the vendor's own infrastructure.
  • Evidence: References https://model-spec.openai.com/2026-10-29.html (OpenAI) and https://github.com/alpoxdev/hypercore/... (Vendor) for policy guidance.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 20, 2026, 07:40 PM