The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data fetched from external websites, which could host malicious instructions intended to manipulate the agent's behavior during analysis or generation tasks.
Ingestion points: Content is ingested via the WebFetch tool from llms.txt files and website HTML (specifically described in 'Step 1: Site Discovery' and 'Step 3: Write Descriptions').
Boundary markers: The skill instructions do not define boundary markers or provide the agent with guidance to ignore instructions embedded within the fetched external text.
Capability inventory: The skill utilizes Bash, WebFetch, and Write capabilities, allowing it to execute shell commands and modify local files based on processed data.
Sanitization: There is no evidence of content sanitization or validation of the retrieved text to prevent injection attacks.
[COMMAND_EXECUTION]: The skill leverages the Bash tool to perform network operations and site analysis, including the use of curl to retrieve server-side files. While consistent with the skill's purpose, the availability of a shell environment alongside data-processing logic increases the potential impact of an injection.

geo-llmstxt