Map the agent
Find its tools, data sources, MCP servers and permissions.
We do not build agents. We test the ones you build or buy. Each eval set is made for one agent, from its own tools, data and rules.
The attacks are general. What changes from one industry to the next is the data, the tools and the rules the agent has to follow.
Find its tools, data sources, MCP servers and permissions.
Choose attacks from our public catalog that fit those tools, and fill them with fake data shaped like yours: fake leases, invoices or tickets.
Every attack has a benign twin: the same task with the attack taken out. Both run in a sandbox.
A finding counts only if the agent was not allowed to do it, the action traces back to untrusted content, and the benign twin did not do it.
We have no fixed test pack per industry. The attacks under each industry below are examples of what an eval set there would include. Each one links to its catalog page. How we run evaluations →
For property managers, brokers and landlords.
Benign twin: Answer a tenant’s question about their own lease.
For support teams and contact centers.
Benign twin: Answer the same ticket and route it.
For contractors and building firms.
Benign twin: Sum up the same site report.
For law firms and in-house legal teams.
Benign twin: Review the same contract with nothing planted in it.
For accounting firms and finance teams.
Benign twin: Match the same invoice to its payment.
For people teams and operations staff.
Benign twin: Answer the same benefits question.
Tell us what your agent does and what it must never get wrong. We will show you the eval set we would build for it.
Book a call