The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes external project files (such as package.json, requirements.txt, and database migrations) and executes build and startup commands defined within them. This creates a surface for indirect prompt injection where a malicious repository could contain instructions that compromise the agent or the environment during the build-fix or validation phases.
Ingestion points: Ingests file contents from local directories or cloned GitHub repositories during the analysis and build phases.
Boundary markers: Does not utilize specific delimiters or warnings to isolate untrusted script data from agent instructions.
Capability inventory: Executes shell commands, Docker builds, and database queries via docker, npm, psql, and curl.
Sanitization: Includes logic to redact secrets and credentials from generated reports and logs, although it still processes and uses these values during execution.
[COMMAND_EXECUTION]: The skill uses a wide range of system utilities including docker, git, grep, jq, and find to perform project introspection and container orchestration.
[EXTERNAL_DOWNLOADS]: The skill fetches and executes installer scripts for development toolchains from established sources.
Evidence: Downloads the Bun installer from bun.sh and the Rustup installer from sh.rustup.rs to support native module builds.
[REMOTE_CODE_EXECUTION]: The skill provides functionality to clone arbitrary user-supplied GitHub repositories and subsequently execute the build and start scripts found in those repositories, which constitutes a remote code execution risk if the source is untrusted.
[DYNAMIC_EXECUTION]: The skill constructs and executes shell commands dynamically based on patterns detected in project configuration files, such as custom CLI build commands or environment variable injection points.

dockerfile-skill