blend0.

Security

What it stops, and what it doesn't yet.

Blend0 sits in the path of an agent's network access and payments. You should know exactly what that protects you from before you rely on it.

01

Non-custodial

Blend0 never holds keys, never signs and never moves money. It decides, and the agent's own wallet signs. There's no pool of funds to steal from Blend0, and nothing for it to lose on your behalf.

02

What it protects against

RiskHow Blend0 handles it
Runaway spend from a loopDaily budget, a cap per request, and budget held for approved payments
One agent spending everyone's budgetPer-agent keys and budgets; unknown keys are refused once agents are set up
Paying an overpriced servicePer-request cap and a threshold above which a person decides
Paying an unknown or failing serviceUnknown services wait for a person; services that fail lose trust
Payment on an unexpected chain or currencyAllowed networks and currencies
Reusing an approvalApprovals work once, for one service and one agent, and expire
Forging or stretching an approvalApprovals are signed; wallet code can check the terms before it signs
An agent approving its own paymentApprovals need an admin token the agent doesn't have
The wrong person approving paymentsRoles (viewer, approver, admin), and on Enterprise, single sign-on that maps your identity provider's groups to roles
A stolen session or a forged requestSessions in HttpOnly cookies with only a hash stored, a CSRF token on every change, and limits on sign-in attempts
Changes nobody can account forEvery sign-in and admin change is in the audit log, with who made it
Reaching hosts it shouldn'tDeny list checked before anything is sent: cloud metadata, internal hosts, anything you name
A friendly name that points at a blocked addressThe address a name resolves to is checked against blocked ranges, cloud metadata by default
A rule change that breaks something or opens a holeAny change can be tested on recorded traffic first, and every change is a version you can restore
An agent gaming the advisor into trusting a serviceSuggestions come only from people's decisions and paid-call results, never from how often an agent asks. Nothing less strict applies without an admin's confirmation
An automatic change that blocks too muchAutomatic changes are off by default, only ever stricter, limited per day, watched, and undone on their own past the limit you set
Disputes about what happenedAppend-only audit log with the reason for every decision

These hold when every outbound call goes through Blend0. See the first limit below.

03

Known limits

Routing is up to you

An agent that calls a service directly, or signs a payment without asking, goes around Blend0. Run agents where their only way out is through Blend0 (the repository has a Docker recipe), and keep the wallet behind code that checks Blend0's signed approval before signing. Wallets that enforce this themselves are on the roadmap.

Keep the admin token from the agent

Approving a payment needs a token Blend0 makes on first run and keeps in a file only you can read. An agent that can read your home folder could read it too, so run agents without access to it.

Keep it private

The admin console shows hosts, prices and decisions to anyone signed in. Keep it on the local machine, or put it on its own port that only your team can reach.

No MFA for password accounts yet

Use single sign-on (Enterprise), so your identity provider enforces multi-factor sign-in.

Addresses are checked once

Blend0 checks the address a name resolves to before sending. DNS can change between that check and the request itself, so a hostile name could still slip through. Keep agents on a locked-down network as well.

Proxy mode sees hosts, not payments

Through the proxy, HTTPS is a sealed tunnel. Blend0 can allow or block the host, but can't see prices inside. Proxy mode is off by default; payments must go through Blend0's fetch.

The log isn't tamper-evident

Anyone with access to the file can change it. Copy it to append-only storage if that matters to you.

Replays are close, not exact

Testing a change replays recorded traffic. It counts spend as it happened, doesn't look up DNS again, and doesn't count payments still waiting. Every test lists these limits; a shadow test on new traffic covers the gap.

04

Report a problem

Write to contact@acificlabs.com with "Blend0 security" in the subject, and steps to reproduce. We reply within three business days. Please don't report security issues in public.

Next: the roadmap