The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data for analysis, creating a potential surface for indirect prompt injection.
Ingestion points: As outlined in SKILL.md, the agent is prompted to accept and analyze various materials like social media posts, news articles, and policy documents provided by the user.
Boundary markers: The instructions lack explicit delimitation or warnings to ignore instructions that might be contained within the analysis materials.
Capability inventory: The skill possesses Bash, Read, and Write capabilities (defined in frontmatter), which could be misused if a processed document contains malicious instructions.
Sanitization: No input filtering or sanitization mechanisms are implemented for the ingested text.
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of scripts to perform quantitative analysis.
Evidence: In the '降級方案' section of SKILL.md, the instructions recommend using Python libraries such as pandas and krippendorff for data processing. This behavior involves dynamic generation and execution of code via the Bash tool to produce frequency statistics.

tw-research-content-analysis