The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes functionality to scan external 'inspiration accounts' on platforms like X and LinkedIn using tools like social-fetch and agent-browser. This process ingests untrusted third-party content into the agent's context to extract structural patterns. While the skill instructs the agent to focus on structure and avoid mimicking phrasing, the lack of explicit boundary markers or sanitization for this external content creates a surface for indirect prompt injection attacks where an attacker-controlled post could attempt to influence the agent's behavior.
Ingestion points: Recent posts from external accounts pulled via social-fetch or agent-browser (referenced in SKILL.md and references/inspiration.md).
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat the scraped content as data rather than instructions.
Capability inventory: The agent has the ability to create and syndicate drafts to a user's Typefully workspace using tools like mcp__typefully__typefully_create_draft.
Sanitization: The skill lacks automated sanitization or filtering of the ingested social media content.

jab-hook