AgentFort security starts with where it runs
Start with the deployment model. Enforcement runs in your environment, next to your services: the gateway that checks agent calls, and the broker that is a sandboxed agent's only way out. The console and control plane run in your own cloud or network. Hosted SaaS is not yet available.
Discovery starts with read-only connectors to systems you already use. The optional sensor for hosts and Kubernetes pods stays inside your environment.
How to verify AgentFort in your environment
Recorded demonstrations introduce AgentFort's workflows. A guided evaluation in your own cloud verifies the integrations, policies and outcomes that matter to your organization.
For a design-partner evaluation, agree the scope and expected results before deploying into a non-production environment in your own cloud:
- Record the AgentFort version, configuration, connected services and policies under test.
- Use synthetic records to check both legitimate actions and actions that should be denied, including access to another customer's record.
- Exercise approval and budget rules, and inspect the downstream system to confirm which actions actually executed.
- Check the data returned to the agent and the audit records. Compare those records with the downstream system's records.
- Test failure behavior and document gaps before deciding whether to expand the deployment.
These are evaluation criteria to agree with your team, not a claim of an independently certified or production-validated release. See the design partner program.
What each control handles and keeps
| Control | What it handles | What it keeps, and what it never keeps |
|---|---|---|
| Discover | Evidence from read-only connectors, cloud logs and the optional sensor | Stores that an agent exists and what it connects to; never prompts, answers or the data agents handle |
| Identity | Proof of who an agent is | Uses the cloud or identity-provider identity an agent already has; with a cloud identity, the agent's secret never leaves the agent |
| Access | Each call: identity, operation, arguments and the record it touches | Service credentials stay in the gateway, scoped to one operation; fields an agent may not see are hidden |
| Data protection | Prompts, files and tool results, in both directions | The language classifier sees text with every value replaced; model providers receive placeholders for personal values |
| Isolation | A sandboxed task's outbound requests, through the broker | Credentials are added outside the sandbox and never enter it |
| Visibility | A receipt for every request | Identity, owner grant, policy and outcome; data protection receipts never hold the values found |
Credentials stay out of agents
Service credentials live in the AgentFort gateway or the Isolation broker, never in agents or sandboxes. The gateway calls each service with its own credential, scoped to the single operation being performed. If that credential leaked, it could do only that one thing. Agents prove who they are with their own identity, not with a shared service key.
Personal data and model providers
Secrets and sign-in material never reach a model provider or anyone outside the company. When an agent asks a model for help, personal values such as an email address, phone number or customer ID travel as placeholders, and AgentFort puts the real values back into the answer. The language classifier that judges meaning only sees text with every value replaced. Data protection receipts record what was found, where, what was done and how long it took, never the values themselves.
Failure behavior and evidence
AgentFort fails closed. If the control plane is unreachable, the gateway keeps the last signed rules for a short lease, 60 seconds by default and configurable, and then refuses every call. Nothing falls back to unchecked access.
Receipts are chained, so a missing or edited entry shows. They are kept on your side and reconciled to the console, and gaps in delivery are reported.
What we can share during a security review
Security reviewers usually want the details before a deployment. Write to [email protected] to:
- Request an architecture and data-flow walkthrough for your environment.
- Send us your security questionnaire for written answers.
For the vendor-neutral view of the risks these controls address, see our AI agent security guide. For who we are, see About AgentFort.
Reporting a vulnerability
If you find a security issue in AgentFort or agentfort.net, email [email protected]. The same contact is published at /.well-known/security.txt, in the format defined by RFC 9116. A good report helps us move quickly:
- Describe the issue and where you found it.
- Include the steps to reproduce it.
- Tell us how to reach you.
We acknowledge every report.