The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill presents a significant attack surface for indirect prompt injection as it ingests feedback from numerous external and attacker-controllable channels. * Ingestion points: Feedback data is collected from public sources such as G2/Capterra reviews, social media mentions (Twitter, Reddit, LinkedIn), and direct customer communications like emails and support tickets (SKILL.md). * Boundary markers: The skill instructions do not specify the use of delimiters, boundary markers, or explicit instructions to the agent to ignore potential commands embedded within the customer feedback. * Capability inventory: The skill utilizes filesystem write capabilities to save reports and chains other tools for review scraping (SKILL.md). * Sanitization: There is no mention of filtering, validation, or sanitization of the input data before it is analyzed by the LLM.
[DATA_EXFILTRATION]: The skill processes and aggregates highly sensitive internal information, including Slack messages, internal email threads, and support ticket details. * The normalized data and final reports contain combined customer signals that may include personally identifiable information (PII) and confidential business strategies. * The reports are saved to a local directory path ('clients//customer-success/voc/'), requiring the user to ensure that this output location has restricted access to prevent unauthorized data exposure.

voice-of-customer-synthesizer