The Agent Skills Directory

[PROMPT_INJECTION]: The scoring_prompt in SKILL.md uses direct interpolation for the {lead_data} variable. This allows an external actor, such as a lead submitting a form, to potentially influence the AI's scoring and routing behavior through malicious content in CRM fields.
Ingestion points: The {lead_data} placeholder within the ai_scoring section of SKILL.md is the entry point for untrusted lead data.
Boundary markers: No delimiters or explicit instructions to ignore embedded commands are present to separate the system instructions from the processed lead data.
Capability inventory: The skill possesses significant write capabilities via the hubspot_assign_owner and salesforce_route tools, along with Slack notification and workflow management.
Sanitization: There is no evidence of data sanitization or validation routines to filter out malicious instructional language from the lead data before it is interpolated into the prompt.

lead-routing