The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The database reference for BRENDA (references/brenda.md) instructs the agent to generate and execute local Python scripts using the third-party 'zeep' library. This technique involves dynamic code creation and execution, which significantly expands the agent's attack surface if the script logic is influenced by malicious input.
[COMMAND_EXECUTION]: Instructions in SKILL.md and multiple reference files (e.g., references/opentargets.md, references/rummageo.md, references/tcga-gdc.md) direct the agent to use shell tools like 'curl' and 'bash' to perform API operations and execute commands. This grants the agent a capability tier that can be exploited if coupled with other vulnerabilities.
[PROMPT_INJECTION]: The skill exhibits a high surface area for indirect prompt injection (Category 8) because it is designed to ingest and return raw data from 78 disparate external public sources. * Ingestion points: API responses from all databases listed in the references folder. * Boundary markers: Absent; the workflow does not specify the use of delimiters or 'ignore embedded instructions' warnings to protect the agent's primary task from data-resident instructions. * Capability inventory: Access to shell tools (curl, bash) and script execution (Python), which increases the potential impact of an injection. * Sanitization: Absent; no mechanisms are described for validating, escaping, or filtering the JSON results before returning them to the user or processing them further.
[EXTERNAL_DOWNLOADS]: Several database references (e.g., references/alphafold.md, references/ena.md) involve downloading scientific data files (such as .pdb or FASTA sequences) from external servers. While the current sources are well-known institutions, the process of downloading and potentially parsing complex file formats adds to the security profile.

database-lookup