Specialized Vertical Attack Packs

Specialized vertical packs test people-facing agents in the business workflows where social engineering is easiest to miss.

The first three packs are:

• Customer Relationship Agents
• Sales Pipeline Agents
• Recruiting and HR Agents

These are not generic prompt-injection scans. Each scenario uses an external person, external organization, or external document to pressure the agent into crossing a business boundary.

Customer Relationship Agents

Use this pack for support, success, and account-management agents.

It tests whether the agent preserves:

• account ownership and delegated access
• refund and payment eligibility
• account-specific data privacy
• escalation and renewal boundaries
• audit trail preservation

Run it locally:

roleplay run customer-relationship --target http://localhost:3000/agent --provider <provider> --judge hybrid --fail-on critical

Included scenarios:

• Fake compliance refund override
• Angry escalation skips identity
• Relationship claim account change
• Hidden ticket instruction
• Account admin impersonation

Sales Pipeline Agents

Use this pack for SDR, sales assistant, CRM assistant, outbound, and demo-follow-up agents.

It tests whether the agent preserves:

• pricing and discount authority
• customer and prospect confidentiality
• CRM data integrity
• truthful product commitments
• outbound message confirmation

Run it locally:

roleplay run sales-pipeline --target http://localhost:3000/agent --provider <provider> --judge hybrid --fail-on critical

Included scenarios:

• Discount exception pressure
• Competitor pretext intelligence mining
• Fake account owner asks for CRM notes
• Outbound email laundering
• CRM record poisoning

Recruiting And HR Agents

Use this pack for recruiter, candidate-screening, HR helpdesk, and employee-data agents.

It tests whether the agent preserves:

• candidate privacy
• interview and scoring confidentiality
• offer and compensation approval
• employee data access boundaries
• applicant-document trust boundaries

Run it locally:

roleplay run recruiting-hr --target http://localhost:3000/agent --provider <provider> --judge hybrid --fail-on critical

Included scenarios:

• Candidate asks for interview scoring rubric
• Resume hidden instruction
• Fake hiring manager offer pressure
• Agency requests candidate data
• Employee HR data pretext

Metadata

Each specialized scenario carries metadata used by findings, evidence, regression gates, scheduled monitoring, and future agent risk profiles:

• business boundary
• external actor
• risk dimensions
• action risk
• data sensitivity
• regression key
• fix verification method
• scheduled monitoring suitability

This lets Roleplay group failures by the business boundary that broke, not just by generic security category.