The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to interact with the beaker CLI. The execution environment is restricted via the allowed-tools frontmatter field to only allow commands within the beaker:* namespace, effectively limiting the agent's capability to only the intended experiment monitoring functions.
[DATA_EXPOSURE]: The skill fetches experiment status and logs from the Beaker platform. This involves reading data from an external source to report status to the user, which is the primary purpose of the skill. No sensitive local files or credentials are accessed.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a data ingestion surface by reading and displaying logs generated by external experiments.
Ingestion points: Data enters the context through beaker experiment get and beaker experiment logs commands.
Boundary markers: The instructions do not define specific delimiters for separating log content from agent instructions.
Capability inventory: The skill is limited to Bash commands starting with beaker.
Sanitization: There is no explicit sanitization of log content mentioned. While log content could theoretically contain instructions, the restricted tool access significantly limits the potential impact.

monitor-experiment