The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill attempts to read project-specific context files such as .agents/product-marketing.md or .claude/product-marketing.md in SKILL.md. This is a local file read operation limited to specific filenames intended to provide business context to the agent and does not involve accessing sensitive system credentials or private keys.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided data, customer feedback, and local marketing context files. Every Category 8 evaluation requires checking for ingestion points, boundaries, capabilities, and sanitization. Ingestion points: user research data and local marketing docs. Boundary markers: not explicitly defined. Capability inventory: no subprocess calls, file writes, or network exfiltration tools are used by the skill. Sanitization: not implemented. While this creates a potential surface for indirect instructions, the lack of dangerous capabilities effectively prevents this from becoming a high-risk attack vector.

ab-testing