Security you can check.
A governance product shouldn’t ask for blind faith. Every control here is one you can see in the architecture and confirm in the signed audit log. No agent behavior to believe, just evidence you hold. Verify, don’t trust.
Six things that are true on every run.
Enforced at the layer below the agent, server-side and test-backed, not policy the agent could talk its way around.
No-route sandbox
Each agent runs in its own container on an internal-only network with no route to the internet. It can’t make a direct outbound connection. There’s nothing to detect because nothing can leave.
DOCKER · NO EGRESSSole-egress broker
One broker is the only door out. It holds your keys and injects them at call time, so the secret never enters the agent. Every call is policy-checked: known model, budget, kill-switch.
BYOK · KEY STAYS OUTEncrypted key vault
BYOK keys are encrypted at rest with per-tenant libsodium sealed boxes, never logged, and never transmitted for our own use. Decrypted in-memory for one call, then gone.
LIBSODIUM · SEALEDSigned audit chain
Every action is Ed25519-signed and hash-chained. Export it and re-verify every event offline on your own machine. Tamper one byte and verification fails.
TAMPER-EVIDENT · OFFLINERBAC + separation of duties
Owner / reviewer / viewer roles enforced server-side. A read-only viewer can’t touch shared keys or billing, and destructive actions are gated by role, so authority is split, not assumed.
ROLE · GATEDBudget kill-switch
Per-run, per-plan, and per-tenant cost caps are enforced in-path. Cross a cap and a host-authoritative watchdog SIGKILLs the run. An escaped agent can’t stop the thing that kills it.
HOST · WATCHDOGDon’t trust the agent. Verify it. Export any run and re-check every Ed25519 signature and hash offline, with no connection to us. The proof is yours to hold. That’s the difference between “secure” and provably secure.
Where we stand, no asterisks. The signed chain is the evidence backbone mapped to NIST 800-171 and CMMC controls, with OSCAL/eMASS export. CMMC L1/L2 is self-assessed; the SOC 2 Type II audit is targeted for Q3 2026. We tell you what’s shipped versus on the roadmap, because a trust product that overstates isn’t one.
Found something? Tell us.
We’d rather hear it from you than from an incident. Responsible disclosure is welcome, and our compliance posture is documented in plain language.
Responsible disclosure
If you find a vulnerability, email security@perimeterone.ai. We read every report, respond, and work the fix in good faith. Please give us a reasonable window before going public.
CONTACT · SECURITY@Trust & Compliance
Where we stand on NIST 800-171, CMMC, and SOC 2 (what’s mapped, what’s self-assessed, and what’s on the roadmap), laid out without asterisks.
Hand the export
to your auditor.
Run a governed agent, export the signed evidence, and let your security team (or an Authorizing Official) re-verify it themselves.