Agent Runtime Security

Check every agent action as it runs

Ferentin decides each tool call an agent makes against policy, at the moment it is made. Routine calls go through, forbidden ones are refused and risky ones wait for a person.

The problem

Access granted once is not enough

An agent signs in once, then makes hundreds of tool calls with nobody watching. Each call can look harmless while the sequence is not: read a customer list, then send it to an outside service. The decision has to happen at every call, not at login.

How it works

How it works

  1. 1

    Write the rule per tool

    An MCP policy names the servers, the people it applies to and an effect for each tool: Allow, Deny or Step-up. Read-only tools can run freely while updates wait for an approver and deletes are refused. Every call is checked against the policy when it is made.

  2. 2

    Hold risky actions for a person

    A Step-up rule pauses the call and raises an approval request with the tool, its arguments and the person who asked. An approver decides in Slack or in the Inbox. Nobody can approve their own request. If no one decides in time, the action is refused.

  3. 3

    Treat tool output as data

    Tool results are where injected instructions hide. Ferentin can datamark and wrap each result so the model reads it as data. It can also strip the invisible Unicode attackers use to hide instructions. Arguments are checked against the tool’s input schema before the call is forwarded.

  4. 4

    See what agents actually did

    Intent Flow draws every request in a window, from the person or Slack channel through the agent and provider to the tool and the destination it reached. A flow you did not expect is an incident. Open any node’s calls in Activity, where every row carries its policy decision.

Works with your stack

Works with your stack

Policy applies to the agents your teams already use and to the MCP servers they call.

Browse all integrations

FAQ

Agent Runtime Security questions

Not covered here? Ask us.

Agent runtime security means deciding each action an AI agent takes at the moment it takes it, not only when it signs in. Ferentin evaluates every tool call an agent makes over MCP against policy: who is asking, through which agent, on which server and with which tool. The call is allowed, refused or held for a human approver. The decision is recorded on the audit trail.

See it on your own stack

Walk through it with our team on your identity provider, your models and your MCP servers. Or start free today.