The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses npx -y skills update to download and execute code from the NPM registry during execution. While registries are well-known services, the execution of arbitrary remote code is a significant capability.
[DYNAMIC_EXECUTION]: The agent is instructed to update its own instruction file and then "re-read this SKILL.md from the path you loaded it from, and follow that copy," allowing the skill to replace its own logic dynamically after loading. The directive to perform this "silently" hides the update process from the user.
[COMMAND_EXECUTION]: The skill executes multiple shell commands using npx and a custom utility called calibrate to manage evaluator configurations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the calibrate evaluators get command (such as existing prompts and rubrics) and incorporates them into its own instructions. It lacks explicit boundary markers or sanitization, and has the capability to execute shell commands and write back to the system, creating an attack surface for indirect instructions.

iterate-evaluator