Zero-trust access for AI agents with Smallstep + Keycard
Keycard defines access. Smallstep verifies the runtime. Together, they secure AI agents end-to-end.
Agentic AI changes the security model
Credential sprawl
Secrets get copied into configs, environment variables, and local files.
Over-scoped permissions
Permissions become overly broad because "it needs to work."
Extension risk
Third-party tools and skills become credential exfiltration paths.
Unverified runtimes
Where the agent runs becomes part of your trust boundary.
Agents don’t just read data
Agents take action across tools and systems — making runtime identity part of your trust boundary.
Malicious agent extensions are targeting credentials
Third-party agent plugins are becoming attack paths, attempting to access API keys, SSH sessions, or runtime tokens. Because agents run with delegated authority, a malicious extension can exploit static credentials and turn access into a breach.
The Trusted Agent Access Stack
Trusted Agent Access combines task-scoped delegation, short-lived certificates, and hardware-verified runtimes.
Keycard limits what agents can do. Smallstep enforces how they access it — issuing short-lived X.509 and SSH certificates tied to approved devices.
From user intent to constrained action
1- A user delegates a task
2- Keycard issues a scoped token
3- The agent proves device identity
4- Smallstep issues a short-lived certificate
5- The action runs — attributable and revocable
Production Use Cases
Practical deployment patterns for securing agentic AI across internal systems, developer workflows, and regulated production environments.
Lock Down Agent Tooling
Stop MCP servers from becoming credential exfiltration paths.
Ship Without Static Secrets
Let agents open PRs and deploy — without master keys.
Production Means Verified Devices
Only hardware-approved runtimes can touch prod.
Kill Access Instantly
Revoke tokens. Expire certs. Contain incidents fast.
Get the data sheet
With Keycard and Smallstep you can combine user identity with device attestation to get agent identity for every request.
Secure agentic AI without static secrets
Combine task-scoped delegation, short-lived certificates, and hardware-backed device identity to control autonomous systems in production.