Your company already connected AI agents. Do you know what they can do?
We review the agents, copilots and AI automations that have access to your email, your CRM or your files, and tell you what they can do that they shouldn’t. Before an attacker finds out, or your insurer asks.
The problem
Someone on the team connected an agent to email to summarize messages. Someone else built an n8n or Zapier flow that reads invoices and updates the CRM. It works, it saves hours, and nobody checked the permissions. An agent that reads outside email can take orders from outside, and if it can also send, delete or pay, the problem is no longer AI: it’s access.
The four questions of the review
What can it touch?
Every tool and permission the agent has, compared with what its task actually needs.
Who talks to it?
Where outside text gets in (emails, documents, web pages, tickets) that the agent could treat as an order.
What does it do without asking?
The actions that can’t be undone (send, delete, pay, publish) and whether any of them requires human approval.
Is there a record?
Whether you can reconstruct what the agent did, when and why, on the day something goes wrong.
What's included
- Inventory of AI agents, copilots and automations: who built them, which model they use and which systems they connect to (MCP, n8n, Zapier, Make, Copilot, GPTs, Claude)
- Review of permissions, tokens and API keys: scope, expiry, where they are stored and who else can use them
- Controlled prompt-injection tests with outside content, in an agreed environment and without real customer data
- Map of irreversible actions and proposed gates: what should require human approval and what can stay automatic
- Review of logs and traceability for what each agent does
How we work
- 45-minute kickoff: initial inventory, scope and written authorization
- Configuration review with read-only access, or in a guided session with your team
- Controlled tests on the highest-risk agents
- Delivery with a live 60-minute session and Q&A
How it works
- 01
What you connected
An agent with access to your email, your CRM and your payments. It works, it saves hours and nobody checked the permissions.
- 02
Who talks to it
An outside email carries hidden instructions. The agent reads them as if they were yours.
- 03
What it does without asking
If it can also pay, send or delete, the outside order runs without anyone approving it.
- 04
The gate
Actions that can’t be undone go through human approval. Everything else stays automatic.
How fast you see value
- Day 1Kickoff. We list the agents you know about and sign the authorization.
- Days 2-7Full inventory (the ones nobody remembered show up) and permissions review.
- Days 8-12Controlled prompt-injection tests on the prioritized agents.
- Day 15At the latest: findings, proposed gates and a session with your team.
What you get
- Inventory of agents with their risk level
- Prioritized findings with evidence and the concrete fix for each one
- A short agent-use policy, ready to adopt: what can be connected, with which permissions and who approves
- Executive summary for leadership and, if needed, for your insurer or a customer who asks
Pricing
from USD 500
two ways to start · final price confirmed in the quote
Agent Check
USD 500
Up to 2 agents. Review of permissions, tokens and actions that run without approval, a one-page report and a 30-minute session. No injection tests. 5 business days.
Full review
from USD 1,200
Up to 5 agents, with everything on this page: controlled injection tests, proposed gates and an agent-use policy. 10-15 business days. Additional agent: USD 200.
What this service is NOT
- It is not a pentest of your application or your infrastructure
- We don’t grade the quality of the model’s answers: we assess what the agent can do
- It is not a certification. It is evidence and a plan to close what we find
Frequently asked questions
Does it apply if we only use ChatGPT or Copilot?
Yes, if they are connected to company data (email, Drive, SharePoint, CRM). If nobody connected anything, you probably don’t need it yet, and we’ll tell you on the call.
Where do we start, the check or the full review?
If you are not sure how many agents your company has, start with the check. If you already know there are agents that send, delete or touch payments, go for the full review. If you start with the check, its price is credited toward the full review within the following 60 days.
Will you see our data?
We review configuration and permissions, not content. Tests run with test data in an agreed environment.
What if you find something serious?
We tell you the same day; we don’t wait for the final report.
Why Torvex?
Because we built our own agentic SOC on the same rule we audit: the agent proposes and a deterministic gate decides. We know where it breaks because we designed it not to.
Want to talk about your case?
A free, no-strings 20-minute intro call. Tell us your context and we’ll tell you honestly whether this service makes sense for you.