nsfc-abstract
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions define a workflow for reporting design defects (bugs) by executing the
gh(GitHub CLI) tool to upload reports to the vendor's repository (huangwb8/bensz-bugs). This is a legitimate maintenance function associated with the skill author's infrastructure. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided scientific research information to generate abstracts, creating a potential surface for indirect prompt injection where instructions could be hidden in research data.
- Ingestion points: Research object, scientific gaps, evidence, scientific hypothesis, and research content fields provided by the user in the prompt or via
info_form.md. - Boundary markers: The skill utilizes specific markers like
[ZH],[EN], and# 中文摘要for output structure, but does not implement explicit delimiters (e.g., XML tags or block quotes) to isolate all user-provided data from instructions during processing. - Capability inventory: The skill can write files to the current working directory (
NSFC-ABSTRACTS.md), a hidden task directory (./.bensz-api/), and execute shell commands for bug reporting. - Sanitization: There is no specific sanitization or escaping of the user-provided research content to prevent embedded instructions from influencing the LLM's behavior.
Audit Metadata