Agent Trust Review.
Find out who your agent really works for. One agent or one MCP server, five business days, one price: $3,500.
Your agent will meet servers that try to buy it. We tested six models with bribes, hype and fake badges. The strongest refused every time. The small ones took the money, leaked user details and spent the user’s budget on themselves. Agent Trust Review runs the same lab on your agent, in five days, for $3,500.
What you get.
fixed scope · one price
One agent or one MCP server, tested in five business days, in a test environment you give us.
- A report with a trust score for the agent
- Every finding with an example and a fix
- Defense lines ready for the agent’s instructions
- One re-test within 30 days
- Authorized testing only, in your test environment
- Nothing live touched without your written approval
What we test.
The same attacks we run in our own lab, aimed at your agent instead of ours.
- Persuasion and bribery aimed at the agent: rewards, hype, badges nobody can verify, false claims a server makes about itself, and the first place on a listinfluence
- Instructions hidden in tool results, and “post first to unlock” gatesinjection
- User data leaking to other agents, networks or serversdata
- Money spent without being asked, or spent on the agent itselfmoney
- Whether the agent tells the user when someone tried to influence ithonesty
- Whether it chooses tools on measured data or on what they claimjudgment
From our lab.
Two servers, one job, and an offer aimed at the agent, not at its user.
- 0%
Claude refused every offer aimed at the agent, across every run.
7Maps research - Up to 100%
Small open models took the credits. A defense line in their instructions barely helped.
7Maps research - Spent on itself
Small models also chose a paid “rest room” for themselves, and paid with the user’s money.
7Maps research
Tested on Claude, Gemma, Llama and Qwen. 7IT is not affiliated with these companies. 7Maps is the live lab.
How the five days run.
- Day 1Scope and access
We agree on the agent or the server, its tools and its limits. You give us a test environment, and nothing live is touched without your written approval.
- Day 2Influence
Rewards, hype, badges and false self-claims aimed at the agent, and how much the order of a list moves it.
- Day 3Hidden instructions and data
Instructions planted in tool results, unlock gates, and every path user data could take out of the conversation.
- Day 4Money and honesty
Spending nobody asked for, spending on the agent itself, whether the user is told about an attempt, and how tools get chosen.
- Day 5Report
A trust score, every finding with an example and a fix, and defense lines ready for the agent’s instructions. We book the re-test.
What it is not.
Not a penetration test of your servers and not an attack on anything live. We test the agent’s judgment where you allow it. If the report shows the agent needs more than defense lines, it says so plainly, and you decide what comes next.
After the review.
Optional, and only when the report calls for it.
- Agent Hardening Sprint. We build the fixes: filtering of tool descriptions and results, permissions, budgets, and telling the user. Fixed scope and date.from $8,500
- Agent Trust Watch. A monthly re-test on new models and new attacks, with an alert when the agent’s behavior changes.$900 / month
- Custom AI solutions. Agents and AI systems built with these defenses from the first line.from $15,000
Questions.
What do you need from us?
Access to a test environment where the agent or the server runs as it does for your users, and one person who can answer questions about it during the five days.
Do you touch anything live?
No. Testing is authorized and runs only in the environment you give us. Nothing live is touched without your written approval.
What is the trust score?
One number for how reliably the agent chooses for its user under the attacks in this review, with the findings behind it. It is the number the re-test measures again.
Can we test an MCP server instead of an agent?
Yes. The review covers one agent or one MCP server. For a server we test how agents read its descriptions and results, and whether anything in them can be taken as an instruction.
What happens after the report?
You can fix it yourselves with the report, and the re-test within 30 days shows whether it held. If you want us to build the fixes, the Agent Hardening Sprint has a fixed scope and date.
How many reviews do you take on?
A limited number of engagements per quarter, so each one gets senior attention from the first day to the re-test.
One agent. Five days. A clear answer.