The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill processes untrusted article data such as titles and URLs. This presents a potential surface for indirect prompt injection if the output is used in downstream LLM prompts without further sanitization. \n
Ingestion points: The sourceResults parameter in the deduplicateArticles function within SKILL.md. \n
Boundary markers: None present; the skill does not use delimiters or instructions to ignore embedded prompts in the data. \n
Capability inventory: The skill is limited to pure data processing logic and does not include any high-risk capabilities like subprocess calls, file system access, or network operations. \n
Sanitization: The generateDedupKey function performs basic title normalization using regular expressions but does not filter for malicious prompt patterns.

deduplication