The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides examples of a 'ReAct Agent' in both SKILL.md and references/examples.md that define a tool for calculations using lambda expr: eval(expr). The input for this function is taken directly from the model's generated output (lm["action_input"]), which is controlled by the prompt and user-supplied data. This enables the execution of arbitrary Python code if the model is successfully tricked via prompt injection.\n- [DYNAMIC_EXECUTION]: The usage of eval() to process and execute strings at runtime is explicitly shown as a recommended pattern for extending the agent's capabilities in the documentation examples.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, creating an attack surface for instructions embedded within that data to be executed by the model.\n
Ingestion points: Functions like extract_entities(lm, text) and react_agent(lm, question) process untrusted strings directly into the model's generation context.\n
Boundary markers: The code examples provided do not utilize delimiters or specific instructions to isolate untrusted data from the model's system instructions, which increases the risk of the model obeying commands hidden in the input.\n
Capability inventory: The skill enables multi-step workflows and specifically demonstrates the use of high-risk capabilities like the eval()-based calculator tool.\n
Sanitization: No input validation, escaping, or filtering of the external data is demonstrated in the provided usage patterns.\n- [EXTERNAL_DOWNLOADS]: The skill references the installation of well-known libraries such as guidance and transformers from standard package registries.

guidance