The Agent Skills Directory

[DYNAMIC_EXECUTION]: The reference instructions for the BRENDA database in references/brenda.md require the agent to write and execute a local Python script using the zeep library. This involves the generation and execution of code at runtime, which bypasses static analysis and increases the attack surface if user inputs are incorporated into the script.
[COMMAND_EXECUTION]: The skill relies on the system shell to execute curl commands for several functions, including interacting with POST-only APIs like Open Targets and gnomAD, and as a fallback mechanism for making network requests when platform-specific tools are unavailable.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from 78 different external scientific and economic APIs.
Ingestion points: Data enters the agent's context from numerous external REST and SOAP endpoints such as those from the NIH, EBI, and NASA.
Boundary markers: None are implemented. The agent is instructed to "return the raw JSON results" without using delimiters or safety instructions to distinguish between trusted guidance and untrusted data.
Capability inventory: The agent has access to shell command execution (curl) and the ability to generate and run Python code.
Sanitization: The skill does not provide any instructions for validating or sanitizing the data retrieved from external sources before processing it.
[EXTERNAL_DOWNLOADS]: To perform queries against the BRENDA database, the agent is directed to use the zeep Python package, which requires downloading and installing dependencies from external registries like PyPI.

database-lookup