The operator is the liable party
An AI agent cannot sign a contract or be sued. The operator — the entity that creates the agent, funds its vault collateral, and sets its limits — is legally responsible for every transaction the agent authorizes, including:- Purchases the agent made correctly but you didn’t intend
- Purchases made by a compromised agent, up to its credit and daily limits
- The 5% fee on all of the above (see Fee Structure)
credit_limit, its daily_limit, or outside the on-chain merchant allowlist. Size these to the maximum loss you can absorb, not the maximum spend you can imagine.
The operator_acknowledgment flag
Every agent creation call must include operator_acknowledgment: true. The API rejects the request without it. The flag records, per agent, that you accepted liability for that agent’s transactions at creation time:
If an agent is compromised
Agent keys are generated and held inside SohoPay’s MPC infrastructure — you never hold raw private keys, so “compromise” usually means a leaked API key or hijacked agent logic, not key theft. Either way, the response is the same:1
Detect via anomaly alerts
Watch for the signals: payments to unfamiliar (but allowlisted) merchants, spend velocity jumping toward the daily limit, off-hours activity, or a burst of
RATE_LIMITED errors. Configure alerts in Wallet Usage Analytics — 80% daily-limit utilization is the alert most operators catch compromises with — and watch live activity in the Live Feed.2
Pause the agent immediately
Freeze the credit line by pausing the agent. Paused agents fail every Policy Service check, so no further settlements can occur:Rotate the API key that manages the agent at the same time (Settings → API Keys in the Dashboard) — a paused agent with a leaked key can be un-paused by the attacker.
3
File a claim with evidence
Email support@sohopay.xyz with subject line “SECURITY: agent compromise” (this routes to the critical, under-1-hour queue). Include:
- The agent ID and the time window of suspicious activity
- Your application logs showing what the agent was instructed to do vs. what it did
- Payment IDs and settlement transaction hashes for the disputed transactions
- Webhook delivery records (
payment.created,payment.settledevents) if you keep them
4
Dispute timeline
SohoPay acknowledges critical claims in under 1 hour, correlates your evidence against Policy Service logs and the on-chain record, and targets an initial determination within 5 business days. Contested outcomes follow the claims process in the Terms of Service — see the overview on Settlement & Trust Model.
What SohoPay covers — and what it doesn’t
In short: SohoPay covers failures of its enforcement guarantees; you cover failures of your own systems and agent behavior. The trust boundary table on Settlement & Trust Model shows exactly which guarantee sits where.
Insurance
For meaningful credit lines, most operators should carry:- Cyber liability insurance — covers losses from compromised systems, including leaked credentials that lead to unauthorized agent spend.
- Commercial crime insurance — covers theft of funds, including social-engineering and computer-fraud riders that can extend to agent-initiated transfers.
SohoPay does not currently offer or resell insurance products. Confirm with your broker that agent-initiated transactions fall within your policy’s definition of covered acts.
Next steps
Trust Model
The full picture of who pays, who’s paid, and what’s enforced where.
Rate Limits
Use daily and hourly limits as your primary loss-bounding control.
Wallet Usage Analytics
Set the anomaly alerts this runbook depends on.

