The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes external text from untrusted sources (URLs, PDFs, arXiv) which can be used to perform Indirect Prompt Injection attacks.
Ingestion points: The source argument in summarize_paper (__init__.py) facilitates the retrieval of untrusted data from external domains or uploaded documents.
Boundary markers: The tool does not implement delimiters or 'ignore' instructions to isolate processed text from the agent's primary instructions.
Capability inventory: The skill returns up to 20,000 characters of external content directly into the agent's context, providing a large surface for malicious instructions.
Sanitization: There is no evidence of sanitization or filtering of the ingested content to remove potential injection vectors.
[DATA_EXFILTRATION]: The tool design permits access to local file paths, which presents a data exposure risk.
Evidence: In __init__.py, the summarize_paper function accepts a source parameter defined in documentation as a 'Path to PDF'. This capability could be exploited to attempt reading sensitive files on the host system if the underlying reader implementation does not strictly validate the file type or location.

paper-summarizer