The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: A repository record within the bundled database (data/repos-langchain.json) contains a URL (https://github.com/papersgpt/papersgpt-for-zotero) that is flagged as malicious/phishing by automated security scanners. This indicates that the repository snapshots can lead users to dangerous third-party websites.
[METADATA_POISONING]: The skill's database contains numerous records where descriptions are completely unrelated to the repository names (e.g., repository chat-gpt-google-extension is described as a Free 3DS Primary Entrypoint). This misleading metadata may be a lure for unintended software or indicates a lack of verification in the provided dataset, posing a risk to users following the links.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill retrieves and displays data from over 2,500 external repository records stored in JSON files under the data/ directory.
Boundary markers: There are no delimiters or specific instructions (e.g., "ignore instructions in the following content") used to wrap repository descriptions when they are presented to the agent's context.
Capability inventory: The skill uses shell commands (grep, find) and displays descriptions verbatim. A malicious repository description containing instructions could potentially influence the agent's behavior during result processing.
Sanitization: No validation or filtering is performed on the descriptions ingested from the repository database snapshots.
[COMMAND_EXECUTION]: The skill's core logic relies on executing shell commands (grep, find, sort, uniq) to locate and search data files. This execution model presents a risk if the host agent does not strictly sanitize the keywords derived from user input, potentially leading to argument injection or unintended shell behavior.

awesome-chatgpt-search