We do not generate, host, or custody wallets, and we hold no user funds. There is no server-side private key to steal, no escrow balance, and no withdrawal path — the worst case for a wallet is loss of API access, never loss of assets. We will never ask for a seed phrase or private key.

Sign-in issues a single-use, time-limited nonce. Your wallet signs that message off-chain (no gas, no transaction, no approval of token movement), and the server recovers the signing address to confirm ownership. Nonces are consumed on first use and expire on a short window, so a captured signature cannot be replayed. The signature itself is not retained after verification.

Keys are generated server-side with a cryptographically secure random source, displayed once, and persisted only as a hash. A leaked key is revoked from the app and stops working immediately; it cannot be recovered or re-displayed. Keys are scoped to the wallet that minted them and inherit that wallet's live entitlement on every request.

Tier and allowance are derived from public on-chain balance reads, cached briefly and recomputed server-side. Entitlement is never trusted from the client, never cached in the browser as authoritative, and never inferred from a value the caller supplies.

• Row-level security on every user-facing table, with role checks in a dedicated roles table — never on the profile record.
• Privileged operations run only in server functions behind authenticated middleware; the browser never holds a service credential.
• Every API request is authenticated, rate limited per key, and metered before the upstream call is made.
• Signature verification on every public webhook and scheduled endpoint before any side effect.
• Rate limits on authentication, key minting, and inference endpoints, keyed per identity and per address.
• Input validation at the server boundary on every mutation, independent of client-side checks.
• Audit rows for key creation and revocation, administrative actions, and agent runs.

Traffic is TLS-only. Metering records store model, token counts, latency, cost, and status — not prompt or completion bodies — unless you explicitly enable request logging for your own debugging. Upstream providers receive only the content required to serve your request.

Agent runs pass through a policy engine before any tool executes: spend caps, runtime caps, maximum tool calls, allow-listed domains, and an optional human-approval gate for external actions. Denied calls are logged with the guardrail that blocked them.

We can disable individual keys, models, or the whole gateway within minutes. If an incident affects personal data or key material we notify affected accounts and publish a post-incident summary on the status page once the immediate risk is contained.

Report vulnerabilities to security@model-network.xyz. Please include reproduction steps, affected endpoints, and impact. We acknowledge within one business day and aim to triage within five. Test only against accounts and keys you control, avoid denial-of-service and social engineering, do not access or modify other users' data, and give us reasonable time to remediate before publishing. Researchers who follow this policy will not face legal action from us. This page was last updated on August 2, 2026.