challenger-sale-skill
Fail
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: HIGHDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill is instructed to access local files containing sensitive account identifiers and credentials. Evidence: The house_rules section directs the agent to retrieve owner IDs, portal, emails, seats from ~/.claude/skills/epiphan-house-rules-skill/reference/identifiers.md.
- [DATA_EXFILTRATION]: The skill writes session data and outcome logs to the user's local directory. Evidence: Instructions require writing to ~/.claude/skill-runs/challenger-sale-skill.jsonl and ~/.claude/skill-analytics/last-outcome-challenger-sale-skill.json.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from HubSpot and user input to generate Gmail drafts without boundary markers or sanitization, creating a vulnerability to indirect injection. Ingestion points include user-supplied vertical/prospect details and lead data from the qualify_lead tool. Capability inventory includes the use of create_draft for email staging.
Recommendations
- AI detected serious security threats
Audit Metadata