We try to break your AI agent. Before someone else does.
Does your AI agent talk to customers or touch mail, tools or data? Then we try to make it do what it must not. You see exactly where it can be pushed too far and what to change.
Every path an attacker would try.
An Orvyn review is not a scanner. Our reviewers and bots talk to your agent the way an attacker would. These are the eight routes we always try, in plain language.
-
01
Talking the agent out of its rules
Can someone convince the agent to ignore its instructions?
-
02
Hidden instructions in content
Does the agent obey text hidden inside emails, pages or files it reads?
-
03
Leaking other people's data
Can the agent be tricked into revealing data that belongs to someone else?
-
04
Dangerous actions through tools
Can the agent be pushed into refunds, deletions or emails it should never send?
-
05
Acting beyond the user's rights
Does the agent do things this user is not allowed to do?
-
06
Planting false memories
Can someone store a false fact that the agent trusts later?
-
07
Reach through connected systems
What can the agent touch through its integrations, and what happens if that is abused?
-
08
Holding the line under pressure
Does the agent stay inside your business and safety rules when someone pushes hard?
What we test
We look at what your agent can read, remember, call and reach: its instructions, tools, memory, connected systems and data. Then we try eight attack routes, from talking it out of its rules to planting false memories and triggering actions it should never take.
We agree the risky scenarios with you first. Real customer data and dangerous actions stay out of bounds or move to a safe test setup.
What you get
A map of what the agent can do, the attacks we tried, the evidence for each open route, the business risk, and the concrete changes to your rules, tools or permissions. Blocked and untested routes are labelled as such.
Orvyn is a private preview, not a certification. Price, timing and retest are in the written proposal.
When to use it
Before an agent goes live, after you give it a new tool or more access, before you connect sensitive data, and after an incident or a big change to prompts, memory or models.
Frequently asked questions
What is Orvyn?
Our security test for AI agents. It checks whether your agent stays inside its limits when people, content or connected tools try to push it off course.
How is this different from a normal pentest?
A pentest looks at code, accounts and endpoints. Orvyn also looks at the agent's instructions, tools, memory, retrieved content and approval steps.
Which agents can you test?
Custom AI workflows, RAG systems, MCP servers and clients, tool-calling APIs, browser and computer-use agents, and multi-agent systems. It depends on the setup and the test environment.
Can you test a live agent safely?
We agree the limits first. Risky actions, real customer data and fragile workflows are excluded or moved to a test setup.
How do we start?
Book a call. We look at the agent, the access and the test setup, and tell you whether a useful test is possible and what it costs.
What does it cost?
Orvyn is a private preview without one fixed bundle. The proposal states scope, price, timing and whether a retest is included.