The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external and potentially untrusted sources to rebuild context, presenting a vulnerability to indirect prompt injection.
Ingestion points: Processes content from file-based or database transcript stores, issue trackers, source control, and external chat/issue channels (referenced in step 4).
Boundary markers: The instructions lack specific delimiting markers or directives to ignore instructions that might be embedded within the processed external records.
Capability inventory: The skill utilizes tools such as git, gh, and ls, and has the ability to spawn subagents for corpus exploration, all of which could be manipulated via injected prompts.
Sanitization: While the skill advises sanitizing output for privacy, it does not include measures for sanitizing or validating ingested data to prevent the agent from acting on instructions found within the logs.
[COMMAND_EXECUTION]: The skill executes shell commands using ls, git, and gh as part of its verification process. The parameters for these commands are influenced by the results of the automated mining of external data.

recall