The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a code example for AutoGen's UserProxyAgent where human_input_mode is set to "NEVER" and code_execution_config is enabled. This pattern allows an AI agent to execute code on the local system autonomously without human-in-the-loop verification, which significantly increases the risk if the agent is manipulated via prompt injection.
[REMOTE_CODE_EXECUTION]: In the CalculatorTool example, the skill demonstrates using the Python eval() function to process the expression input string. Using eval() on strings that may originate from untrusted sources (such as LLM outputs or external data) is a critical security vulnerability that allows for arbitrary code execution within the host environment.
[PROMPT_INJECTION]: As the skill describes building agents that integrate tools, memory, and multi-agent communication, it inherently establishes an attack surface for indirect prompt injection. If an agent built using these patterns processes untrusted external data (e.g., via the search tools or memory retrieval described), it could be coerced into executing malicious commands or tools.

agent-development