The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to run build or test scripts defined in a hackathon team's local repository clone. These repositories are provided by external participants, making them untrusted sources. Executing these scripts (e.g., via npm test or a repository's documented build process) allows for arbitrary code execution on the host machine.
[INDIRECT_PROMPT_INJECTION]: The agent is tasked with reading chat threads and repository code from hackathon teams. This untrusted content could contain adversarial instructions designed to influence the agent's actions, such as manipulating drafts or misusing the hackathon CLI.
Ingestion points: Chat threads via hackathon chats read and repository files.
Boundary markers: None present to delimit untrusted content.
Capability inventory: Shell execution via hackathon CLI and git, repository build/test execution, and file reading.
Sanitization: The instructions recommend avoiding certain types of scripts (e.g., deployment scripts), but there is no technical enforcement or validation of the content.
[COMMAND_EXECUTION]: The skill uses the hackathon CLI tool and git for managing platform requests and repository interactions. Actions include claiming requests, posting replies, and synchronizing code.
[DATA_EXFILTRATION]: The skill reads specific files from the team's repository clone based on names found in chat threads. This provides a mechanism to access files that might contain secrets or sensitive information if the agent is misled by instructions found in the untrusted external content.

hackathon-coach-help