The Agent Skills Directory

[DATA_EXFILTRATION]: The skill requires access to sensitive communication data, including CRM records and email 'Sent' folders, to perform voice extraction and pattern analysis for the organization (documented in references/improve-outreach.md). While this involves reading private data, the skill does not exfiltrate this information to external endpoints, instead using it to generate localized outreach drafts for user review.
[PROMPT_INJECTION]: The skill contains logic allowing the agent to dynamically update its own instructions. It is directed to modify its reference Markdown files (SKILL.md and reference files) to record new principles, do/don't rules, and organizational preferences derived from user feedback or pattern extraction.
[PROMPT_INJECTION]: The skill operates on an attack surface susceptible to indirect prompt injection due to the ingestion of untrusted external data.
Ingestion points: Untrusted data enters the context through LinkedIn engagement (posts and comments), website visitor activity logs, third-party intent data (e.g., G2, Bombora), and incoming email/CRM threads.
Boundary markers: There are no defined boundary markers or isolation instructions to separate untrusted external content from the agent's internal instruction-processing logic.
Capability inventory: The agent can read comprehensive communication history, draft outreach messages for multiple channels, and modify its own reference files to store new operating rules.
Sanitization: The skill does not implement explicit sanitization or filtering of external data, relying entirely on the 'Stage for manual approval' mandate as the primary defense against malicious influence on the final output.

reach-out