The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents and encourages the use of load_from_pickle and save_as_pickle for data handling in SKILL.md and references/importing-exporting.md. Python's pickle module is inherently unsafe as it can be used to execute arbitrary code during deserialization. If an attacker provides a malicious .pkl file to the agent, it could lead to complete system compromise.
[INDIRECT_PROMPT_INJECTION]: The skill defines a significant ingestion surface for untrusted external data in SKILL.md and references/importing-exporting.md.
Ingestion points: The functions load_from_mgf, load_from_mzml, load_from_msp, load_from_json, and load_from_usi ingest text, XML, and JSON-based spectral data.
Boundary markers: The instructions do not define boundary markers or instruct the agent to ignore instructions embedded in the spectral metadata or peak comments.
Capability inventory: The agent has the capability to write files (save_as_mgf, save_as_json), potentially execute code (load_from_pickle), and perform network operations (load_from_usi).
Sanitization: There is no evidence of sanitization or validation of the input strings to prevent prompt injection attacks targeting the LLM context.
[COMMAND_EXECUTION]: The skill instructions in SKILL.md provide command-line installation instructions using uv pip install, which is a standard procedure but involves executing shell commands to manage environment dependencies.

matchms