Obsign sits in the path of every action your AI agents take. It decides allow / deny / redact in microseconds — before the action runs — and seals a cryptographically tamper-evident, court-grade record of that decision that an examiner can verify offline. Even your own operator can't rewrite history.
Agents act autonomously through tool calls — approving, paying, trading. You need two things at the same instant: to stop an out-of-policy action before it executes, and to prove, court-grade, exactly what was decided and why. Today these are separate and breakable — policy gateways write editable logs, and audit trails are mutable by the very operator who runs them. When a regulator asks “prove your agent didn't move that money,” most teams can't.
At the Model Context Protocol boundary — the emerging standard wiring agents to their tools — every action is decided synchronously and sealed asynchronously, off the hot path.
In-process, formally-verified policy — allow, deny, or redact — before the action executes. Fail-closed by construction: an error never becomes an allow.
Hash-chain + Merkle tree + signed checkpoint + an independent witness. Editing, deleting, reordering, or back-dating any record is detectable — even by the operator.
A self-contained evidence bundle anyone can re-check with only a public key — no trust in Obsign required. Sensitive inputs are hashed, never stored.
A US regulator already endorsed the exact design — a tamper-evident audit trail.
The SEC's 2022 amendment to Rule 17a-4 explicitly permits a tamper-evident audit trail as an alternative to write-once storage. FINRA's 2026 priorities ask firms to keep audit trails, require human checkpoints before execution, and block out-of-bounds AI actions. Obsign is built to that standard.
Built test-first against an adversary that continuously tries to forge the seal. The demo blocks a live $4.8M agent wire inline, seals it, and an offline verifier rejects a back-dated copy.
If your agents take consequential actions, you'll be asked to prove they stayed in policy. We're working with a small number of design partners. Let's talk.
Request a pilot