skills/yushui2022/liuhuan-mathmodel-skills/modeling-paper-rubric-and-model-selector/Gen Agent Trust Hub
modeling-paper-rubric-and-model-selector
Warn
Audited by Gen Agent Trust Hub on May 14, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill is designed to ingest external text (
paper_prompt_text) or files (paper_prompt_path) and treat them as 'strong constraints' (强约束). There are no defined boundary markers or sanitization logic, allowing these external inputs to potentially override the agent's primary safety or operational instructions via indirect prompt injection. - [COMMAND_EXECUTION]: The skill instructs the agent to read content from arbitrary filesystem paths provided by the user via the
paper_prompt_pathparameter. This capability can be exploited to read sensitive system files, configuration, or credentials if the agent's environment does not strictly sandbox file access. - [DATA_EXFILTRATION]: By reading files from user-supplied paths and 'incorporating' their contents into the generated paper output or summaries, the skill creates a vector for data exposure. Content from the filesystem is processed by the LLM and reflected in the user-visible output, which could lead to the unintentional display of private data.
Audit Metadata