The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to autonomously run shell commands and benchmark scripts as part of its "Evaluator" phase. A reference example in references/evaluator-patterns.md specifically instructs the agent to use sudo nginx -s reload, which poses a significant risk of unauthorized privilege escalation.
[REMOTE_CODE_EXECUTION]: The core pattern involves the agent modifying a "Sandbox" file and then executing it. The provided Python template in references/evaluator-patterns.md uses importlib.util.module_from_spec and spec.loader.exec_module(mod) to dynamically load and run the modified code at runtime, creating a framework for autonomous arbitrary code execution.
[DATA_EXFILTRATION]: The evaluator patterns include network-capable tools such as wrk for load testing and the anthropic library for LLM-based scoring. Combined with the agent's ability to read arbitrary files to populate the "Sandbox," these tools provide a mechanism for exfiltrating sensitive data from the environment during an optimization run.
[PROMPT_INJECTION]: The skill has a high surface for Indirect Prompt Injection. The agent is instructed to read results.tsv (ingestion point) to analyze history and generate hypotheses. This file is written by the agent based on evaluator output, which could be poisoned by the code being optimized. No boundary markers or sanitization logic is provided to prevent instructions in the data from influencing the agent's logic (Category 8).

autoresearch