Don’t trust the agent.
Verify it.

The industry is racing to give AI agents real power: your repo, your shell, your APIs, your customers’ data. The tooling to contain that power, and to prove what an agent actually did, has lagged badly behind. Most “AI security” asks you to trust a model’s behavior, or a vendor’s dashboard. We don’t think that’s good enough for anyone who has to answer for what their agents do.

So we built P1 Halo from the other direction. Each agent runs in a sandbox with no route to the internet. A single broker is the only door out. It holds your keys, injects them at call time so they never enter the container, and policy-checks every call. Every prompt, tool call, and output signs into an Ed25519 hash-chained audit log you can re-verify offline, on your own machine, with no network and no trust in us. Tamper with one byte and verification fails.

That’s the whole idea: governance you don’t have to take on faith. We map the evidence to NIST 800-171 and CMMC controls so the same signed chain that protects a solo developer’s run can stand up in front of an auditor or an Authorizing Official. We’re plain about scope, too: CMMC is self-assessed, and SOC 2 Type II is targeted for Q3 2026, not yet attested.

PerimeterOne LLC is a real, registered, veteran-owned company, founder-built by someone who has carried responsibility for systems that had to hold up under scrutiny, with a background spanning security and government-contracting work. That history shapes the product: when you’ve had to answer for what a system did, you stop accepting “trust the dashboard” as an answer. We hold ourselves to the same bar we sell: we run our own governed agents in production, behind the same broker, the same sandbox, and the same signed chain we ship to you. We’re our own first customer, and we’d rather earn trust by handing you proof than asking for it.