The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains multiple examples of 'ReAct Agent' implementations that use Python's eval() function to execute tool logic based on model-generated input.\n
Evidence: In SKILL.md (Pattern 4) and references/examples.md (Agent Systems and Tool Use with Validation), the calculator tool is defined as lambda expr: eval(expr).\n
Risk: The input passed to eval() is generated by the LLM via the gen('action_input') command. If the model is manipulated to produce malicious code instead of a mathematical expression, that code will be executed with the permissions of the agent process.\n- [EXTERNAL_DOWNLOADS]: The documentation instructs the user to install third-party dependencies from standard package registries.\n
Evidence: The skill references pip install guidance, transformers, and llama_cpp.\n
Context: These are well-known open-source libraries provided by reputable organizations and are considered safe for the intended use of this skill.\n- [PROMPT_INJECTION]: The skill includes patterns that are vulnerable to indirect prompt injection due to a lack of input sanitization and boundary enforcement.\n
Ingestion points: The skill ingests untrusted text data through the question argument in the react_agent function and the text argument in extract_entities.\n
Boundary markers: The implementation lacks specific markers (such as XML tags or explicit instructions) to prevent the LLM from following commands embedded within the processed data.\n
Capability inventory: The agent has high-impact capabilities, specifically the ability to execute code via the eval() tool function.\n
Sanitization: There is no validation or sanitization logic applied to the generated action_input before it is processed by the execution tool.

guidance