The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute multiple Python scripts via the Bash tool, including scripts/generate_schematic.py, scripts/compliance_checker.py, scripts/budget_calculator.py, and scripts/deadline_tracker.py. None of these scripts are provided within the skill's file list. Instructing an agent to execute unverified code that is not part of the distribution package represents a significant security risk.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted research descriptions provided by users to generate grant proposals. As the skill is granted powerful capabilities through the Bash, Write, and Edit tools, the absence of explicit sanitization, validation, or boundary markers (such as XML tags or 'ignore embedded instructions' warnings) for this external data creates a vulnerability to indirect prompt injection attacks.
[PROMPT_INJECTION]: The instructions in SKILL.md utilize forceful and mandatory language (e.g., '⚠️ MANDATORY: Every research grant proposal MUST include...', 'This is not optional') to dictate specific agent behaviors. These patterns act as behavioral overrides that may conflict with the agent's default safety constraints or user-defined preferences.
[COMMAND_EXECUTION]: The skill explicitly authorizes the use of the Bash tool to run local Python scripts and generate assets. The reliance on shell execution for scripts that are not provided or audited increases the attack surface for arbitrary command execution.

research-grants