Multi-party computation (MPC) wallets remove the single private key, but they do not remove the people, devices, policies and software that decide what gets signed. Institutional custodians, exchanges, payment firms and corporate treasuries now hold large balances in threshold signing wallets, where several parties each hold a key share and jointly produce a signature without ever reassembling the key. The cryptography is strong when implemented correctly; most real failures sit in protocol implementation flaws, weak share custody, approval policies that can be bypassed, and signers who approve what a compromised screen shows them. This guide is operational: how to deploy, run and monitor an MPC wallet estate safely. It is for custody, treasury, security and key management teams. For the underlying mathematics, see the Secret Sharing and Threshold Cryptography Guide; for exchange-wide controls, see the Crypto Exchange Identity and Key Security Guide.
Key takeaways
- MPC moves risk, it does not remove it. The attack surface becomes the share holders, the signing nodes, the policy engine and the API identities that request signatures.
- Protocol implementation matters. The 2023 BitForge disclosure showed that missing zero-knowledge proofs in GG18 and GG20 implementations could let an attacker extract a full private key; ask every provider which protocol they use and how it was reviewed.
- The policy engine is the real control. Quorums, value limits, address allowlists and approval rules decide whether a theft succeeds; protect and audit the policy itself as a privileged object.
- Signers need an independent view of the transaction. The Bybit theft in 2025 showed that genuine signatures on a tampered transaction are still a loss, whatever the signing scheme.
- Every share holder is an identity. Human approvers, cosigner servers and API clients each need verified identity, least privilege, separation of duties and fast revocation.
MPC, multisig and HSM compared
MPC threshold signing (often called TSS) splits signing authority across n parties so that any t of them can sign, while the full key never exists in one place, not even at generation if distributed key generation (DKG) is used. On-chain multisig achieves a similar quorum with several independent keys and a smart contract or script that checks them. A hardware security module (HSM) protects a single key inside certified hardware. Many custodians combine them, for example MPC with each share held inside an HSM or secure enclave.
| Property | MPC threshold signing | On-chain multisig | Single key in HSM |
|---|---|---|---|
| Single point of key compromise | No, if shares are on independent infrastructure | No, if keys are on independent devices | Yes, the HSM and its access controls |
| Chain support | Any chain whose signature scheme the protocol supports (ECDSA, EdDSA, Schnorr) | Only chains and wallets that support multisig scripts or contracts | Any chain whose curve the HSM supports |
| On-chain visibility of quorum | None: the chain sees one ordinary signature | Full: signers and threshold are visible and auditable on chain | None |
| Changing signers | Off chain by share refresh or resharing; address can stay the same | On-chain transaction to change owners | Key rotation usually means a new address |
| Fees and privacy | Same as a single-key transaction | Higher fees and visible structure | Same as single key |
| Main assurance question | Is the protocol and its implementation sound and reviewed? | Is the multisig contract and its interface sound? | Is the HSM certified and are its access controls strong? |
| Typical failure | Implementation flaw, colluding or compromised share holders, policy bypass | Signer interface compromise, contract bugs, signer collusion | Stolen operator credentials, over-privileged API access |
Decision guidance: choose MPC when you need many chains, an unchanged address when signers change, and private quorum structure. Choose on-chain multisig when public verifiability of the quorum matters, for example a protocol treasury or DAO, and see the Multisig Treasury Operations Guide. A single HSM key is acceptable only for low-value hot wallets with strict outflow limits. Small treasuries with a few signers often get better assurance from hardware wallets in a well-run multisig than from an MPC service they cannot inspect.
Reference architecture and where shares live
A typical institutional deployment has three to five share holders across different trust domains:
- Provider or cosigner node: a share held by the custody provider or your own server, often inside a secure enclave or HSM, which applies policy before taking part in signing.
- Organisation server share: a share in your own infrastructure, ideally on separate cloud accounts or on premises, with its own administrators.
- Approver device shares: shares on mobile devices or hardware held by named approvers, protected by device secure hardware and biometrics.
- Backup or recovery share: an encrypted share held offline or with an independent party, used only for recovery under a documented ceremony.
The design rule is independence: no single administrator, cloud account, identity provider or software update channel should be able to reach t shares. If all shares run on one cloud tenancy managed by one platform team, the threshold is cosmetic. The CryptoCurrency Security Standard (CCSS v9) makes a similar point for multi-signer wallets: key material must be stored and used on different logical or physical devices (1.05.9.1), stored in different locations (1.02.3.1), and at Level III held by distinct operators or separate entities (1.02.4.1). See the CryptoCurrency Security Standard Guide and the Confidential Computing Guide for enclave-based share protection.
Protocol assurance: what to ask about the cryptography
Threshold ECDSA in particular is complex, and several widely used protocol families have had implementation flaws. In August 2023 Fireblocks researchers disclosed BitForge, affecting more than 15 wallet providers and projects. For GG18 and GG20 implementations, a missing zero-knowledge proof could let a malicious party extract the full private key, with the number of signatures needed ranging from as few as 16 to as many as a billion depending on the implementation. For Lindell17 implementations, mishandling of failed signatures could leak key material over roughly 200 signing attempts. Affected providers patched, but the lesson is that "we use MPC" says nothing about safety on its own.
NIST runs a Multi-Party Threshold Cryptography project. Its First Call for Multi-Party Threshold Schemes, NIST IR 8214C, was finalised on 20 January 2026 and covers threshold signatures, key generation and related primitives; NIST IR 8214B, an initial public draft from August 2022, discusses threshold EdDSA and Schnorr. There is no NIST standard for threshold ECDSA yet, so assurance rests on published protocols, independent review and the provider's own engineering.
| Question for the provider or your engineering team | What a good answer includes |
|---|---|
| Which threshold protocol is used for each signature scheme? | Named protocol and paper, version, and any deviations from the published specification |
| Are all zero-knowledge proofs in the protocol implemented and checked? | Yes, with evidence from an independent cryptographic review |
| How are aborted or failed signing sessions handled? | Failures are logged, parties that cause repeated aborts are identified and blocked |
| How is the key generated? | Distributed key generation, so no party ever sees the full key; or a documented ceremony if generated centrally |
| How are shares refreshed? | Proactive refresh on a schedule and after any suspected compromise, without changing the public key |
| Who reviewed the implementation and when? | Recent independent cryptographic audit, bug bounty, public disclosure process |
| What happens if the provider disappears? | Tested export or recovery path that does not depend on the provider |
Key generation and share ceremonies
Treat share generation, refresh, resharing and recovery as ceremonies with a signed runbook, named roles and evidence. CCSS v9 Level III requires a documented, controlled generation process (1.01.3.2) with a secure environment, equipment checks, a signed runbook, defined roles with segregation of duties and each actor generating their own keys, plus random number generation conforming to NIST SP 800-90A (1.01.3.1). The same discipline applies to MPC:
- Prepare: confirm the participants, devices and software versions; verify software signatures; record who is present.
- Generate: run DKG so each participant creates its own share; record the resulting public key and address.
- Verify: sign and broadcast a small test transaction, then confirm each share holder can take part in signing.
- Back up: create encrypted backup shares according to the recovery design; store them in separate locations with tamper-evident packaging or an independent custodian.
- Record: keep a ceremony log with timestamps, participants, hashes of software used and witness sign-off.
- Refresh: schedule proactive share refresh (for example quarterly) and run it immediately when a share holder leaves or a device is lost.
Use the Cryptographic Key Ceremony Template as a starting point.
The policy engine: what decides a signature
In most MPC platforms, a cosigner only joins a signing session when the transaction passes policy. That policy is where theft is stopped or allowed.
| Policy control | Example rule | What it stops |
|---|---|---|
| Destination allowlist | Withdrawals from cold storage only to listed warm wallets; new addresses need a cooling-off period and two approvals | Direct theft to attacker addresses |
| Value and velocity limits | Per-transaction and 24-hour limits per wallet and per initiator | Large single drains, slow draining under the radar |
| Quorum by risk | Two approvers under a threshold value; three including a senior approver above it | Single insider or single compromised approver |
| Initiator separation | The person or API client creating a transaction cannot approve it | Self-approval by an insider or a stolen API key |
| Contract call restrictions | Only listed contract methods (for example transfer) on listed contracts; block approvals and delegate calls by default | Signing a hidden ownership change or unlimited token approval |
| Time windows | Cold wallet movements only in business hours, with a delay before execution | Out-of-hours attacks, gives time to cancel |
| Policy change control | Policy edits need a separate admin quorum and a delay, and are alerted | Attacker loosening limits before a theft |
Policy changes deserve their own quorum. If one administrator can raise the limit and then approve a withdrawal, the threshold adds nothing.
Where identity fits: approvers, cosigners and API clients
An MPC wallet estate is an identity system. Human approvers, signing servers and the API clients that create transaction requests are all identities with signing power.
- Human approvers: verify identity before granting access, check references and run background checks where law allows, as CCSS 1.05.3 to 1.05.5 require; use phishing-resistant authentication; review approver lists monthly and revoke on the day someone leaves, followed by share refresh.
- Cosigner and node identities: run on hardened, dedicated hosts or enclaves with separate administrators; protect node credentials as privileged non-human identities and monitor their use.
- API clients: treasury systems, payment engines and trading bots that initiate transactions hold API keys or certificates. Scope each to specific wallets and actions, store them in a secrets manager, bind them to source networks and rotate them. A stolen initiator credential should be stopped by the approval policy, never sign alone. See the Secrets Management Guide.
- Administrators: admins of the MPC platform, the identity provider and the cloud accounts hosting shares can often reach several shares indirectly; apply privileged access controls and separate them. See the Privileged Access Management Guide.
How to implement and migrate
- Classify wallets: hot, warm and cold, with value limits and target signing times for each.
- Design the quorum and independence: t-of-n per wallet class, with shares across independent infrastructure, administrators and locations.
- Complete protocol due diligence: answers to the questions above, independent review evidence, and a provider-independent recovery path.
- Write policy as code where supported: allowlists, limits, quorum rules and contract call restrictions, versioned and reviewed.
- Pilot with low value: run test wallets for several weeks, exercise approvals, rejections, share refresh, device loss and full recovery.
- Migrate in tranches: move balances in stages with verification after each, keeping the old wallets locked but recoverable until the new estate is proven.
- Roll back plan: if a protocol flaw or provider issue is disclosed, have a tested procedure to sweep funds to a fresh wallet under a different scheme, such as hardware multisig.
Detection and measurement
Log every signing request, policy decision, approval, rejection, aborted signing session, share refresh, device enrolment and policy change, and ship logs to a system the MPC administrators cannot alter. CCSS 2.02 Log and Monitor sets the expectation for logging and monitoring. Alert on new destination addresses, approvals outside usual hours or locations, repeated aborted signing sessions (which can indicate a malicious party probing the protocol), policy edits and new API clients.
- Policy coverage: share of wallet value under allowlist and limit policies; target 100 percent for warm and cold wallets.
- Time to revoke: time from an approver leaving or losing a device to removal and share refresh.
- Recovery test success: date and result of the last full recovery test per wallet class.
- Independence score: number of distinct administrator groups and infrastructure domains holding shares for each wallet.
- Policy exceptions: count of emergency overrides per quarter, each with a recorded reason.
Framework mapping
| Framework | Reference | Application to MPC wallets |
|---|---|---|
| NIST SP 800-53 Rev. 5 | SC-12, SC-12(1), SC-13 | Key establishment and management, key availability and recovery, approved cryptography |
| NIST SP 800-53 Rev. 5 | AC-5, AC-3, AU-6 | Separation of duties between initiators, approvers and admins; policy enforcement; log review |
| NIST CSF 2.0 | PR.AA-05, PR.DS-01 | Least privilege and separation of duties for signers; protection of shares at rest |
| CCSS v9 | 1.01, 1.02, 1.04, 1.05, 2.02, 2.04 | Key generation, multi-signer design, access, usage, logging, key compromise documentation |
| MITRE ATT&CK | T1552.004 Private Keys | Theft of key material or shares from hosts and backups |
Common failure patterns
- Trusting the signing screen: in the Bybit hack, signers approved a cold wallet transfer through a multisig web interface that attackers had tampered with, and around $1.5 billion was stolen. MPC approvers face the same risk; verify transaction details on an independent device or with policy checks that do not rely on the interface.
- Shares without independence: all shares on one cloud account, managed by one team, through one identity provider.
- Unreviewed protocol code: using an MPC library or provider without evidence that all zero-knowledge proofs are implemented and reviewed, the gap BitForge exposed.
- Weak policy change control: limits and allowlists editable by a single admin with no delay or alert.
- No share refresh after leavers: a departed approver's device still holds a valid share.
- Untested recovery: backup shares that have never been used to restore a wallet, or recovery that depends entirely on the provider.
Practitioner checklist
- Document the threshold protocol, version and independent review for every signature scheme in use.
- Map each share to its infrastructure, administrators and location, and remove any path to t shares through one party.
- Enforce destination allowlists, value limits and initiator separation for all warm and cold wallets.
- Put policy changes under their own quorum with a delay and alerting.
- Verify transaction details on an independent channel before high-value approvals.
- Scope, vault and rotate every API client credential that can initiate transactions.
- Run share refresh on a schedule and on every approver change.
- Test full recovery at least annually, including a provider-independent path.
- Monitor aborted signing sessions, new addresses, policy edits and out-of-pattern approvals.
Standards and references
- NIST: Multi-Party Threshold Cryptography project
- CryptoCurrency Certification Consortium: CCSS v9 requirements table
- Fireblocks: BitForge vulnerabilities in MPC wallet implementations
- NIST: SP 800-53 Rev. 5
- NIST: Cybersecurity Framework 2.0
- MITRE ATT&CK: Unsecured Credentials: Private Keys
Related NHI Mgmt Group resources: Crypto Exchange Identity and Key Security Guide · Secret Sharing and Threshold Cryptography Guide · Multisig Treasury Operations Guide · Crypto Wallet Security Guide