Built to sit in your critical path.
Arc decides whether your AI agents can act. That only works if you can trust it — so the security model is designed to be inspected, not taken on faith.
Default-deny policy
Anything an agent didn't explicitly get permission for is blocked. No surprise capabilities.
Signed execution (ES256)
Approved work runs as a short-lived signed request your app verifies before any side effect. The agent never holds a broad API key.
Hash-chained audit
Every attempt, approval, and block is hashed against the one before it. Re-order or delete a record and the chain breaks. Redacted by default.
Sensitive inputs encrypted at rest
The raw input of a guarded action is stored encrypted at rest with AES-256-GCM (field-level), then decrypted only to deliver it inside the signed execution request. The audit log itself stays redacted.
Verifiable, not taken on faith
Every approved action is a signed (ES256) request your app verifies before any side effect, and every decision lands in a hash-chained audit you can re-check. You trust the guardrail because you can verify its decisions yourself — not because we ask you to.
Hardening & honesty
We don't claim certifications we don't have. The current focus areas on the road to production-grade are documented openly: external immutable audit export, egress controls around signed delivery, key rotation, and a third-party security review. We'd rather tell you what's done and what's next than imply more.
Report a vulnerability
Found something? Please report it privately via our contact page rather than a public issue. We'll acknowledge quickly and keep you posted on the fix.