The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes the JavaScript Function constructor to evaluate arithmetic results produced by the AI model in the Game of 24 benchmark.
Evidence: The function evaluateExpression in SKILL.md contains const result = Function("\"use strict\"; return (${sanitized})")();.
While the code includes a sanitization step using expr.replace(/[^0-9+\-*/().]/g, ''), the use of dynamic execution for content sourced from model outputs is a risky implementation pattern.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data in the form of AI model responses across several benchmarks (MATH, GSM8K, Game of 24).
Ingestion points: The runBenchmark function in SKILL.md collects responses from a generateFn and passes them to evaluators like evaluateGSM8K and evaluateMATH.
Boundary markers: The prompt templates (e.g., cotPrompt, metaPromptGame24) do not use clear delimiters or instructions to prevent the agent from obeying instructions that might be embedded within the benchmark problems themselves.
Capability inventory: The skill has the ability to execute code (via Function) and manipulate string data based on these untrusted inputs.
Sanitization: While basic arithmetic characters are whitelisted for the math evaluator, there is no general sanitization or escaping of the model-generated responses before they are processed by the evaluation logic.

prompt-benchmark