Runtime security by design
Runta enforces a kernel-level isolation boundary around every agent run. Policy is declared, not patched. The enforcement layer is separate from your agent code.
Isolation boundary model
The Runta enforcement perimeter intercepts every syscall, network egress, and resource allocation before it reaches the host. Your agent code runs inside; the policy engine runs outside.
What the enforcement layer covers
These are capability descriptions, not compliance certifications. Each control is enforced at the runtime layer, independently of your agent code.
Resource quotas
Per-run limits on CPU time, memory, wall-clock duration, and LLM token consumption. Limits are enforced at the cgroup level. A runaway agent cannot consume host resources beyond its declared cap.
Egress allowlist enforcement
All outbound network traffic is blocked by default. Only domains and IP ranges you explicitly declare in your policy YAML are reachable. Enforcement is at the network layer, not the application layer, so agent code cannot bypass it.
Append-only audit logs
Every enforcement decision, quota event, and egress attempt is written to an append-only log. Records are hash-chained: each entry includes a cryptographic reference to the previous one. Within the retention window, no record can be modified or removed.
Syscall surface restriction
A seccomp-bpf profile blocks syscalls outside the permitted set for each agent profile. Calls that would allow privilege escalation, container escape, or filesystem traversal beyond the declared mount points are denied before execution.
Secrets injection without exposure
API keys and credentials are injected into the sandbox at runtime via a dedicated secrets channel. They are never written to the filesystem or passed through environment variables visible to process inspection. The agent sees only the resolved value at call time.
Policy version control
Policy YAML is stored with full revision history on the Governed and Enterprise plans. Each run record references the policy version that was active at execution time, giving you a durable link between observed behavior and the declared intent.
See the enforcement model in action
The Sandbox tier is free and exposes every control. Run your first governed agent in under ten minutes.