The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, untrusted data such as grant solicitations (NOFOs) and investigator-authored proposals, creating a surface for prompt injection attacks.
Ingestion points: SKILL.md instructs the agent to "extract the requirements from this NOFO" and "critique my R01 aims", which involves processing content provided by the user or fetched from remote sources.
Boundary markers: The instructions lack explicit requirements for delimiters or boundary markers (such as XML tags or triple quotes) to prevent the agent from mistaking data-embedded instructions for system rules.
Capability inventory: The skill is granted access to Read, Write, Edit, and Bash tools, which could be misused if a malicious NOFO triggers an injection.
Sanitization: There is no evidence of sanitization, filtering, or validation logic applied to the external grant content before it is interpolated into the agent's context.
[DATA_EXFILTRATION]: The skill utilizes the OpenRouter API (openrouter.ai) for optional scientific figure generation. While the skill includes a transparent warning about third-party data transmission, this constitutes a network operation to a domain outside the standard whitelist.
[COMMAND_EXECUTION]: The skill provides instructions to execute a local Python script (skills/scientific-schematics/scripts/generate_schematic.py) to generate project diagrams. This represents a capability to execute shell commands based on Natural Language descriptions provided in the skill logic or by the user.

research-grants