The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied architecture designs and static fixtures, creating a surface for potential indirect prompt injection.
Ingestion points: User-provided material is ingested through the prompt logic defined in SKILL.md and prompts/resilience-testing.md.
Boundary markers: No explicit structural delimiters (such as XML tags or unique markers) are specified to separate user data from system instructions.
Capability inventory: The skill's instructions strictly prohibit high-risk actions, including network access, fault injection, and reading sensitive local credentials.
Sanitization: The agent is required to perform an "Input Audit" to verify data freshness and source, which serves as a mitigation strategy against accepting injected instructions as facts.- [SAFE]: The skill instructions contain explicit prohibitions against dangerous operations, such as calling production systems or accessing real dependencies, ensuring the agent remains in an analytical mode.

resilience-testing