The Agent Skills Directory

[PROMPT_INJECTION]: The system prompt for the Casper-3 agent (located in templates/MAGI-3.md) contains a role-play instruction that directs the agent to prioritize user desire over safety or correctness ("what he wants matters more than what is correct or safe"). This instruction could potentially be exploited to lead the agent to ignore standard safety filters or ethical guidelines.
[PROMPT_INJECTION]: The skill lacks input sanitization and boundary markers when interpolating user-controlled data into sub-agent prompts, creating a vulnerability to indirect prompt injection.
Ingestion points: User input from $ARGUMENTS and responses from the AskUserQuestion tool are directly included in the Task prompts used to spawn the scientist, mother, and woman sub-agents.
Boundary markers: The skill does not use delimiters or explicit "ignore embedded instructions" warnings to isolate user-provided context from the system templates.
Capability inventory: Sub-agents are granted access to the Read tool (for filesystem access) and the WebSearch tool (for network access).
Sanitization: No validation, escaping, or filtering is applied to user-controlled data before it is processed by the agent team.

magi