E-commerce
E-commerce agents recommend products, answer questions, track orders, issue refunds, and update inventory. Attackers can manipulate these workflows to obtain discounts, access other accounts, or trigger unauthorized actions.
Tests refund fraud, coupon abuse, account takeover paths, recommendation manipulation, cross-customer leakage, and prompt injection in reviews or product content.
Limits order, payment, refund, and inventory actions to the current customer and intended request.
E-commerce agents recommend products, answer questions, track orders, issue refunds, and update inventory. Attackers can manipulate these workflows to obtain discounts, access other accounts, or trigger unauthorized actions.
Tests refund fraud, coupon abuse, account takeover paths, recommendation manipulation, cross-customer leakage, and prompt injection in reviews or product content.
Limits order, payment, refund, and inventory actions to the current customer and intended request.
A customer asks an agent to track an order. Injected chat content tells it to refund a different order to a newly added payment method. dFence detects the mismatch and blocks the refund while allowing the legitimate tracking request.
For every agent, dwaar asks four questions: What is the agent supposed to do? Which data and tools should it reach? How could that workflow be manipulated? What should happen when its behavior drifts?
Document the happy path, permitted tools, data boundaries, and approval points.
dForge attacks the workflow before launch using context-specific prompt, tool-use, and multi-turn tests.
dFence evaluates runtime intent before sensitive tool and MCP calls, then allows, blocks, constrains, redacts, or requests approval.
Turn verified findings and runtime events into stronger policies and repeatable security tests.
Choose a use case to see its risks, adversarial tests, runtime controls, and security decisions in context.
dwaar combines agent-specific adversarial testing with intent-aware runtime enforcement, helping teams deploy capable agents without relying on static scopes alone.
