The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data (diff files and fixture directories), which could contain embedded malicious instructions intended to influence the agent's behavior during the evaluation process.
Ingestion points: Diff files, rubrics, and fixture directories provided as inputs to the task (SKILL.md).
Boundary markers: The instructions require adherence to specific problem IDs and rubric questions, which provides a level of structure but does not fully isolate the LLM from the content of the files.
Capability inventory: The agent is instructed to run validation checks and tests on the fixture code, presenting a vector for code execution driven by potentially malicious data.
Sanitization: The skill explicitly directs the agent to review code before execution and isolate it from production data or credentials.
[COMMAND_EXECUTION]: The procedure includes running validation tasks or scripts found within the provided fixture code. While the skill instructs the agent to honor resource limits and isolate the environment, the execution of arbitrary scripts from untrusted fixtures is a security risk characteristic of its primary evaluation purpose.

score-eval