An AI agent connected to your systems is a user with permissions. Test it like one
A security audit for AI agents, chatbots and MCP connections, before they go to production or after they already have.
In short
An AI agent reads emails, documents and databases, calls tools and acts on behalf of the organization. Malicious text in a document or email can make it take an action you did not intend, or disclose data. So we test what the agent can do, not only what it is supposed to do: permissions, tools, data, and actions that need a person's approval.
What we test
- Permissions: which systems the agent is connected to, what it can do in them, and whether it has more access than it needs.
- Prompt injection: attempts to make the agent take unwanted actions through malicious instructions in documents, emails and web pages.
- Data leakage: what the agent can disclose, to whom, and what is kept in its logs and memory.
- MCP connections and tools: who authenticates whom, where the keys are kept, and what happens when an external server changes.
- Irreversible actions: where a person's approval is required (payment, deletion, sending to a customer).
- Monitoring and stopping: whether you can see what the agent did, and whether you can stop it at once.
- Cost: what each action costs, and what happens under real load.
What you get
A findings report by severity, a fix for each finding, and a clear recommendation: go to production, fix before going live or keep running, or stop until fixed. If you want, we can also make the fixes.
The scope and the price are set after a short call about the agent and the systems it is connected to.
Two tracks
- Before production: a pilot that works in a demo, and it is not clear whether it is ready for real users.
- An agent already running: a review of what actually happens, based on the logs and permissions.
Why us
We build RAG systems, chatbots and MCP connections ourselves, and we know where they break. We also run secure development training for developers.
Frequently asked questions
Is this a penetration test?
Not exactly. A traditional penetration test does not necessarily cover how the model responds to malicious instructions or its permissions to tools. We focus both on the agent's behavior and on the security of its connections and permissions.
Do you need access to the code?
It is better, because that way we find more. We can also start with an outside-in review.
The agent runs on a ready-made platform, not on code we developed ourselves. Is the audit relevant?
Yes. On a ready-made platform too, the permissions, the connections and the data the agent can access are important points of risk.
SLAtech LTD