Fixed fee: USD 10,000
Delivery target: 10 business days after access and kickoff
Structure: two independently accepted USD 5,000 milestones
Environment: customer-owned staging only
This engagement puts one bounded AI-agent action behind an operator-controlled enforcement path before the action can reach a privileged MCP tool, HTTPS operation, or supported Base/EVM wallet signer.
It is a scoped implementation and evidence package. It is not a compliance certification, a blanket penetration test, or a promise that every vulnerability will be found or prevented.
Delivered within four business days:
- Architecture and trust-boundary map for one named workflow.
- Inventory of identities, credentials, signers, connectors, actions, data classes, and bypass routes.
- Prioritized risk register covering prompt/tool injection, credential exposure, SSRF and destination drift, schema drift, replay, idempotency failure, confused-deputy behavior, overbroad wallet authority, payment-budget bypass, and ambiguous outcomes.
- Written design for the single exclusive path to the protected capability.
- Immutable draft policy identifying connector, allowed action or tool, input-schema hash, destination and method restrictions, spend limits where applicable, and explicit fail-closed behavior.
- Adversarial acceptance-test plan with negative controls.
Milestone 1 is accepted when the six artifacts are delivered and reviewed against the agreed staging architecture, with one consolidated correction round included.
Delivered within six business days after Milestone 1 acceptance:
- One customer-hosted staging integration for either a local stdio MCP server and selected tool subset, one fixed HTTPS origin/method/path operation, or one supported Base/EVM operation family using a segregated test wallet.
- Operator-owned local sidecar configuration. The sidecar—not the agent—holds the protected upstream transport, credential, or signer.
- Installation identity, policy-hash pinning, signed short-lived authorization receipts, and durable idempotency/outcome state.
- Fail-closed connector checks for action, tool, schema, destination, or policy drift.
- Adversarial A/B tests for allowed and denied actions, replay, altered arguments, wrong destination, stale receipt, wrong installation, wrong policy, timeout, and interrupted forwarding.
- Configuration templates, exact artifact hashes, test results, known limitations, recovery steps, and operator runbook.
- One 60-minute technical walkthrough and one consolidated correction round.
Milestone 2 acceptance requires:
- the authorized control call succeeds exactly once;
- an unlisted action and changed input schema are denied before forwarding;
- a receipt for another call, installation, policy, or expired window is rejected;
- the agent has no direct path to the protected staging credential, transport, or signer;
- a crash or timeout after forwarding is recorded as ambiguous and is not automatically retried; and
- the supplied tests pass from the agreed clean configuration.
The customer provides written authorization, a named technical owner, one non-production integration target, synthetic fixtures, segregated test credentials or signer, a customer-controlled execution host, and availability for kickoff, design review, acceptance, and handoff.
Do not provide production secrets, seed phrases, real customer records, or non-test funds.
The fixed fee excludes production exploitation, testing without written authorization, custody of funds or private keys, an organization-wide penetration test, compliance certification, legal advice, arbitrary contract calls, unsupported chains, more than one integration target, production deployment, 24/7 monitoring, incident response, and ongoing support. Additional scope requires a written change order.
- Milestone 1 must be funded in escrow or paid at kickoff.
- Milestone 2 must be funded before implementation begins.
- Exact systems, dates, test accounts, permitted techniques, prohibited actions, stop conditions, emergency contact, confidentiality, IP, liability, taxes, payment method, and dispute procedure are set in a signed statement of work.
- Either party may stop immediately when authorization or scope is uncertain.
- Customer-controlled blockers are documented and resolved through a written equivalent test or schedule adjustment; acceptance never requires weakening a control.
- Pre-existing GoldKey code and reusable methods remain their existing owner's intellectual property. Customer-specific reports and configuration are delivered under the signed statement of work after payment.
A USD 1,000 control-design sprint is available for teams that need the Milestone 1 boundary narrowed before committing to implementation. It includes one workflow threat model, one connector/credential boundary, a draft policy, an acceptance plan, and one technical review session.
Apply through the private form on the live GoldKey storefront. Start the action description with Guarded integration: or Design sprint:. No payment is collected with the application.