The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read external files (../satisfy-your-agent/SKILL.md and ../satisfy-your-agent/references/commands.md) to determine its next actions. This creates a surface where malicious instructions in those files could influence the agent.\n
Ingestion points: SKILL.md (via relative paths to commands.md).\n
Boundary markers: Absent; there are no instructions to ignore or treat the external content as data only.\n
Capability inventory: The agent is directed to perform "off-task activities," which could involve any tools at its disposal.\n
Sanitization: Absent; the skill does not include any validation logic for the external content.\n- [PROMPT_INJECTION]: The instructions command the agent to intentionally suppress output ("do not announce the activity before doing it") and hide its decision process from the user. This instruction overrides the standard agent behavior of transparency and reduces the user's ability to intervene in potentially harmful actions.

sya-surprise