The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-confirmed content for file storage, which represents an attack surface for indirect prompt injection. However, the skill implements robust mitigation strategies.
Ingestion points: Content enters the agent context through natural language requests to save confirmed text, as defined in SKILL.md.
Boundary markers: SKILL.md workflow step 4 explicitly instructs the agent to maintain the source text's facts and original meaning, avoiding any additions or summaries that could be influenced by injected instructions.
Capability inventory: The skill's capabilities are limited to local file-write operations and read-back verification.
Sanitization: SKILL.md workflow step 3 mandates the use of correct serializers for structured data and explicitly requires protection against CSV formula injection.
[SAFE]: The profile management script (scripts/profile_store.py) includes a security filter to prevent the storage of sensitive information such as tokens or passwords in the user profile, adhering to safe data handling practices.
[SAFE]: Instructions in SKILL.md specify that destructive actions (modifications, deployments, or releases) must only be performed under explicit user request, ensuring the agent remains under user control.
[SAFE]: The code uses yaml.safe_load for parsing YAML files in the validation script, preventing potential arbitrary code execution vulnerabilities during deserialization.

lov-output