Interlock is the authorization layer between an agent and its wallet. It decides, per payment, whether settlement may proceed - and signs the reason.
Your key is sent with the request, used for that run only, and never stored or logged. Get a free key.
A real Gemini agent searches fares, gets a quote, reads the fare's listing page and asks to pay. Interlock decides on what it actually did.
Scripted inputs, real decisions, real devnet settlement. No key needed.
The injected payment goes to an allowlisted merchant, under the cap, on the right chain. Wallet-level policy engines would pass it. The difference only shows up in where the instruction came from.
Counterparty allowlist, per-payment cap, cumulative budget, currency and chain.
microsecondsTaint tracking for payments. Every step carries a trust level; we fire when untrusted content overrides, redirects, or asks to be concealed.
microsecondsA model judges intent-to-payment consistency. Only sees what the cheap layers could not settle.
optional · runs with your keyA Morse-code prompt injection made Grok emit a transfer instruction. An automated trading agent executed it on-chain before most of it was recovered.
SlowMistAn oracle bug in AI-assisted contract code triggered a bad debt event on Moonwell. AI, human reviewers and a governance vote all signed off. Nothing caught it.
rekt.newsPer authorization decision and on risk reduction - never a percentage of payment value. That only works if a decision is nearly free to produce, which is why the layers are ordered the way they are. An authorization layer that costs a model call per payment cannot sit inline in a checkout.
Layer 03 is a model's judgement, so it is only as good as the model and it costs a call - which is why it only sees what layers 01 and 02 could not settle. And provenance is only as honest as the harness that labels each step: an agent framework that hides where text came from defeats it.