Build a crypto custody risk program: key management, MPC vs HSM, withdrawals,…
- User
Build a crypto custody risk program: key management, MPC vs HSM, withdrawals, attestations, and incident response.
Assistant
The user wants a comprehensive Crypto Custody Risk Program. This is a high-level request requiring a structured, professional, and technical response, likely suitable for a CISO, CRO (Chief Risk Officer), or Head of Custody Operations at a crypto firm (exchange, fund, or custodian).
Key components requested:
- Key Management
- MPC vs. HSM (comparison and strategy)
- Withdrawals (operational security)
- Attestations (audit/compliance)
- Incident Response
Target Audience: Senior Management/Technical Risk Officers. Tone: Professional, authoritative, structured, and actionable.
Structure of the response: I need to structure this as a formal policy framework or a strategic roadmap.
- Executive Summary: What is this program and why does it exist?
- Pillar 1: Key Management & Storage Strategy (MPC vs HSM): This addresses the core technology.
- Pillar 2: Withdrawal Operations & Transaction Security: The operational flow.
- Pillar 3: Governance & Attestations: The audit/compliance layer.
- Pillar 4: Incident Response (IR): The "break glass" procedure.
- Summary Checklist: A quick reference.
Drafting Content - Section by Section:
-
1. Key Management (The Foundation):
- Need to define the lifecycle: Generation, Storage, Usage, Rotation, Destruction.
- Key concept: Segregation of Duties.
- Key hierarchy (Hot vs. Warm vs. Cold wallets).
- Self-correction: Don't just say "use cold wallets." Be specific about air-gapped vs. MPC.
-
2. MPC vs. HSM (The Tech Debate):
- This is a major decision point.
- HSM (Hardware Security Module): Traditional, FIPS 140-2/3 compliant, rigid, great for single points or centralized trust, but can be a single point of failure (SPOF) if not configured in HA clusters.
- MPC (Multi-Party Computation): Modern, distributed, flexible, removes single point of failure. Good for operational fluidity.
- Recommendation: A hybrid approach is usually best for large institutions. Use MPC for operational efficiency (warm) and HSM/Air-gapped for deep cold storage. I should provide a comparison table.
-
3. Withdrawals & Transaction Controls:
- This is where the money leaves.
- Controls: Velocity limits, whitelisting (address book), 2FA/MFA.
- Human element: Dual control (4-eyes principle).
- Travel Rule compliance (FATF).
-
4. Attestations & Audits:
- How do we prove we have the money?
- Proof of Reserves (PoR) using Merkle Trees.
- Third-party audits (SOC 1/2 Type II).
- On-chain monitoring (chainalysis/elliptic) for AML.
-
5. Incident Response (IR):
- What happens when (not if) things go wrong?
- Playbooks: Key compromise, smart contract bug, insider threat.
- Crucial addition: The "Emergency Key" or "Recovery Key" procedure.
- Communication plan (transparency).
Refining the "Risk Program" aspect: It's not just a tech list; it's a Risk Program. So, I need to frame it around Risk Management (Identify, Protect, Detect, Respond, Recover). I will add a Governance layer.
Detailed content generation:
- MPC vs HSM Table: I need to compare them based on Security, Usability, Redundancy, and Cost.
- Withdrawal Flow: Draft a visual description of the flow (Initiation -> Validation -> Signing -> Broadcast).
Self-Correction during drafting:
- Initial thought: Just talk about Bitcoin.
- Correction: Must be chain-agnostic or cover multi-chain risks (EVM vs. UTXO).
- Initial thought: Focus heavily on code.
- Correction: This is a risk program, so focus on policy and governance, not just code.
Final Polish: Use formatting (bolding, bullet points, headers) to make it readable. Add a "Golden Rule" or "Core Philosophy" section to tie it all together (e.g., "Assume Breach").
Outline for Output:
- Program Charter & Governance: Roles (Key Ceremony Admins, Approval Committee).
- Technology Stack: MPC vs. HSM: Detailed comparison and recommendation.
- Operational Security: Withdrawals: Limits, Whitelisting, Multi-sig logic.
- Verification & Compliance: Attestations: PoR, SOC2, Chainalysis.
- Incident Response: Playbooks, War rooms.
Let's write the response. To build a robust Crypto Custody Risk Program, you must assume a "Zero Trust" architecture. The goal is not just to prevent theft, but to ensure Survivability (business continuity) and Auditability (proof of reserves).
Here is a comprehensive framework for a Custody Risk Program, structured by critical domains.
I. Governance & Key Management Philosophy
Before selecting technology (MPC vs. HSM), you must define the Key Hierarchy and Policy Engine.
1. Key Hierarchy
- Level 1: Master Seed (Root Key): The genesis of all addresses. Highest risk. Must be generated offline (air-gapped) and sharded (Split using Shamir’s Secret Sharing).
- Level 2: Operational Keys: Derived from the Master Seed. Used for daily hot/warm wallet operations.
- Level 3: API/Session Keys: Used by internal microservices to communicate with node providers or signing servers.
2. Policy Engine
Every transaction must pass through a generic policy engine before signing is even attempted.
- Velocity Limits: Max $X per hour/day.
- Destination Whitelisting: Only allow withdrawals to pre-approved addresses (with a time-lock cooldown for new additions).
- Velocity & Volume Anomalies: Flag transactions 2x the standard deviation of average volume.
II. Technology Selection: MPC vs. HSM
The industry is shifting from hardware-centric security to cryptographic-centric security. Here is the Risk Analysis for choosing between them.
Feature HSM (Hardware Security Module) MPC (Multi-Party Computation) Concept Keys live inside a tamper-resistant physical box. The private key never exists in one place; it is split into mathematical shares. Single Point of Failure High. If the HSM fails or is compromised, keys are lost or stolen. Low. To sign, an attacker must compromise multiple devices/servers simultaneously. Availability Requires physical redundancy (clustering). High. Shares can be distributed globally. Scalability Hardware provisioning is slow/expensive. Software-defined; scales with cloud infrastructure. Chain Support General purpose, but curve support varies. Highly flexible; can add new curves/chains via software updates. Recommendation: The Hybrid Approach
A modern risk program should utilize Threshold MPC for operational efficiency, protected by HSMs for key material storage.
- Strategy: Use MPC to split a key into 3 shares (e.g., Server, Mobile Device, Backup). The "Server" share is encrypted and stored inside an HSM. This combines the cryptographic security of MPC with the physical tamper-resistance of HSMs.
III. Operational Security: Withdrawals & Transaction Signing
This is the "Last Mile" of security. You must prevent unauthorized transactions and "Business Logic Attacks" (e.g., the DAO hack logic).
1. Hot vs. Cold vs. Warm Wallet Architecture
- Hot Wallet (Online): Holds <5% of assets. Insured. Keys are in memory (or HSM connected to network). High velocity.
- Warm Wallet (Online, isolated signing): Holds 15-20%. Used to refill Hot wallets.
- Cold Vault (Air-gapped): Holds >75% of assets.
- The "Data Diode" method: Transaction data flows out via API, but signing happens on an offline machine, and the signed transaction is broadcast via QR code or USB one-way transfer.
2. The Signing Ceremony
For high-value movements (Cold to Warm), enforce a Quorum:
- M-of-N Multi-Signature: Require 3 out of 5 authorized signers.
- Geographic Dispersion: Signers must be in different jurisdictions to prevent physical coercion attacks ($5 wrench attack).
- Video Attestation: Signers must be on a recorded video call during the ceremony.
IV. Assurance & Attestations
In crypto, "Don't Trust, Verify" is literal. You must provide proof to stakeholders (investors, regulators, users).
1. Proof of Reserves (PoR)
- Liabilities vs. Assets: You must prove you hold the private keys to the assets and that your liabilities (user balances) match the assets.
- Merkle Tree Approach: Build a Merkle tree of all user balances. Users can verify their leaf node inclusion without seeing neighbors' data.
- Auditor Role: A third party (e.g., Armanino, Mazars) verifies the root of the tree matches the on-chain balance.
2. On-Chain Monitoring
- Implement a separate monitoring tool (e.g., Chainalysis, TRM Labs) that watches your custodial addresses.
- Alert: If funds move from a Cold Wallet that wasn't initiated by a known signing ceremony, trigger an immediate halt.
V. Incident Response (IR) Plan
Crypto IR is distinct from traditional IT IR because of the immutability of the blockchain. You cannot "delete" the hack; you can only stop the bleeding.
1. The "Kill Switch"
- Every custodian must have a Circuit Breaker.
- If suspicious activity is detected, the system automatically pauses all outgoing transactions.
- This "Pause" functionality must be hard-coded into smart contracts (for tokenized assets) or enforced by the policy engine.
2. The "War Room" Protocol
- T+0 to 15 mins: Activate the IR team. Revoke access for all non-human service accounts.
- T+15 to 60 mins: Assess scope. Is it a key leak? Or a logic bug?
- T+1 Hour: If keys are compromised, initiate Key Rotation (if possible) or migrate funds to a pre-provisioned emergency "safe" wallet.
3. Post-Mortem & Bounty
- Engage blockchain forensics firms (e.g., CipherBlade) to trace stolen funds.
- Consider a "White Hat" bounty offer to the hacker (e.g., "Return 90%, keep 10%").
Summary Checklist for Risk Leadership
- [ ] Key Ceremony: Are seeds generated on air-gapped hardware and sharded?
- [ ] Separation of Duties: Does the developer who wrote the withdrawal code have access to the production keys? (Should be No).
- [ ] Testing: Do we run regular "Fire Drills" moving funds from Cold to Hot to verify key holders know the process?
- [ ] Audit: Is the custody infrastructure SOC 2 Type II compliant?
- [ ] Insurance: Do we have a Specie Insurance policy to cover physical loss of keys or theft?
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.