sci-ssci-polishing

Pass

Audited by Gen Agent Trust Hub on Sep 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides and utilizes a local utility script scripts/check_invariants.py. The agent is instructed in SKILL.md to execute this script to perform a deterministic audit of scientific invariants (numbers, citations, and specific terms) between the source and revised text. This script operates on data passed via a JSON file and uses standard Python regex and counting logic.
  • [INDIRECT_PROMPT_INJECTION]: As a translation and polishing tool, the skill accepts and processes untrusted academic prose, creating a surface for potential indirect prompt injection attacks.
  • Ingestion points: Untrusted Chinese or English academic text ingested via the primary input prompt as described in SKILL.md.
  • Boundary markers: The instructions define a rigorous 'preservation ledger' and an audit workflow (Workflow Steps 2 and 5) to ensure the agent remains faithful to the source material, which acts as a logical barrier to instructions embedded in the manuscript.
  • Capability inventory: The agent has the ability to execute shell commands to run the provided auditing script.
  • Sanitization: The auditing script uses regex to count tokens and does not interpret the input text as executable code or control instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 5, 2026, 02:30 AM
Security Audit — agent-trust-hub — sci-ssci-polishing