Bitcoin Custody: Blockstream's Offline Security Case
See how Blockstream and Thales combine HSMs and MPC, why offline signing matters, and what institutions should assess before adopting cold custody at scale.

Why Blockstream Is Taking Bitcoin Custody Back Offline
Institutional Bitcoin custody is a constant trade-off: you want strong protection for assets, but the people approving transactions still need a system they can actually use. Many custody setups rely on multi-party computation (MPC) to spread signing authority across devices or locations. It can reduce dependence on any single key - but it also brings real complexity. In some designs, parts of the signing workflow still end up depending on "normal" computing environments.
Blockstream's pitch is different. It's building a custody system with French technology group Thales that pairs Blockstream's secure MPC with Thales hardware security modules (HSMs). The stated goal is to keep the full signing protocol inside dedicated security hardware. Blockstream co-founder and CEO Adam Back argues that simpler, offline signing can cut exposure to remote attackers.
That argument deserves a close look, but it's not a universal solution. Cold storage changes the threat model - it doesn't erase operational risk, insider risk, or the need for solid recovery procedures. Below is how the announced system is described, what offline custody can (and can't) protect against, and what institutions should weigh before they pick a setup.
How the Blockstream and Thales system works
Blockstream's announcement pairs two components: its secure-MPC technology and Thales HSMs. An HSM is purpose-built hardware for protecting cryptographic operations. Instead of letting key-related computation happen on general-purpose computers, an institution can use an HSM to control where sensitive operations occur and limit how they're accessed.
MPC (multi-party computation) is a way to generate a signature through cooperation among multiple parties or devices without one complete private key existing in a single place. In custody, that can help distribute authority. The details matter, though. MPC protocols aren't all the same, and implementation choices determine what data and computation move around during signing.
Blockstream says its architecture runs the entire signing protocol inside Thales' HSM environment. That's the core distinction in how the company presents it. Some MPC systems do portions of signing outside dedicated security hardware. In Blockstream's description, the protocol stays within the HSM boundary, while MPC coordinates authorization.
The intent is to mix distributed signing with hardware-enforced protection. But it doesn't mean the broader custody system has no software, no networks, or no human process around it. Operators still have to initiate requests, verify transaction details, manage access, and maintain the devices. The security boundary is narrower than the whole operational environment.
What the design is trying to protect
A custody system has to protect more than a private key "as a secret." It also needs to prevent unauthorized transaction approvals, restrict access to signing devices, and support a controlled process for legitimate transfers. Keeping signing operations inside an HSM can reduce how many places do sensitive computation - depending on the exact setup and deployment.
Still, this is a design goal, not proof that any specific deployment is secure. An institution has to understand how signing requests reach the HSM, who can approve them, what the device exposes to operators, and how recovery works if hardware is lost or damaged.
What is not established by the announcement
The announcement lays out an architecture and the security benefits Blockstream expects from it. It doesn't provide enough detail to compare every operational aspect against other custody products. Buyers should ask for the threat model, independent assessment results, supported workflows, recovery procedures, and evidence of how the HSM boundary is enforced in practice.
That procurement reality matters. A promising architecture doesn't automatically translate into a safe deployment. Security depends on the whole package - hardware, software, people, policies, and incident response - not on the label attached to one component.
Why offline signing changes the threat model
When signing is done in an internet-connected environment, you inherit a remote attack surface. Firewalls and access controls help, but they don't make a connected system magically unreachable. Attackers can target employee devices, credentials, administrative tooling, or other systems that influence the transaction workflow. If those systems can reach the signing infrastructure, compromise elsewhere can become relevant to custody.
Adam Back's argument is that taking signing equipment offline removes a route for remote access. He pointed to an incident involving Binance and relied on the exchange's investigation, which described attackers compromising employee devices and obtaining credentials that allowed access to wallet-signing infrastructure. The takeaway isn't that every connected custodian will fail the same way. It's that the line between everyday workplace systems and signing authority needs to be designed and tested, not assumed.
Cold storage aims to reduce this exposure by keeping signing devices disconnected from routine networks. A transaction can be prepared in one environment, moved to the signing system through a controlled process, approved offline, and then returned to a connected environment for broadcast. Each handoff between environments becomes a control that has to be managed deliberately.
Offline is a control, not a complete security strategy
Air gaps make remote compromise harder, but they don't make a system invulnerable. An attacker might go after people who handle transaction data, exploit a compromised endpoint before the transaction is moved, or abuse an operator's legitimate access. Physical access, tampering, supply-chain risk, and sloppy backups still matter.
Offline workflows can also create their own failure modes. Manual handling can introduce delays or mistakes. If transaction details aren't checked carefully on a trusted display, an operator could approve the wrong destination or amount. Even a device that's disconnected most of the time still needs secure updates, monitoring, and a recovery plan that's actually been tested.
The operational price of reducing connectivity
Custody choices aren't just about security - they're also about whether you can run the system day to day. A cold-signing process may work well for reserves that don't move often, but it's awkward for businesses that need frequent, time-sensitive withdrawals. A slower flow can be an acceptable control for long-term holdings; for funds used in regular settlement, the service delays can become a real problem.
So the question isn't whether offline signing is safer "in the abstract." It's whether the reduced remote attack surface is worth the extra operational burden for a specific pool of assets. Many organizations will need different workflows for long-term reserves and for frequently used liquidity rather than forcing every Bitcoin transaction through the same offline process.
Where MPC complexity creates security trade-offs
MPC is often sold as a way to avoid storing one complete private key in one place. That can be useful - especially when signing authority must be distributed across teams, locations, or devices. But MPC's security properties depend on the protocol and how it's implemented. A system might include multiple parties, multiple rounds of communication, software components, and recovery procedures. Each extra interaction adds more "moving parts" to understand and verify.
Back's criticism targets MPC protocols he sees as too complex and interactive for strict offline cold storage. His point isn't "MPC is bad." It's that an architecture designed for connected, interactive operation may not match a workflow where devices are intentionally offline. If a protocol expects frequent communication, it can clash with strict cold-storage procedures.
Blockstream's response - according to the company - is to run its signing protocol inside Thales HSMs. That's meant to preserve MPC-based authorization while limiting where signing computation happens. Whether it's a better fit still depends on implementation details and on what the organization actually needs. In other words: treat the company's description as a design claim to validate, not a replacement for diligence.
Complexity is a risk factor, not a verdict
Complex systems are harder to review, test, and operate. That doesn't automatically mean they're insecure. A more involved design can be justified if it solves a real requirement - distributed control or a specific recovery model, for example. The important question is whether each layer earns its place and whether the team can explain how it behaves when something goes wrong.
A practical review looks at what happens when one signer is unavailable, an HSM fails, a network is down, or a transaction is submitted incorrectly. It also checks which components must communicate, which parties can approve transfers, and whether the organization can recover without weakening its normal controls. If those answers aren't clear, adding more cryptography won't fix the operational problem.
AI may increase pressure on weak implementations
Back also warned that AI could help attackers find subtle flaws in complex MPC implementations. That's a reasonable concern, but it shouldn't be read as proof that AI has already broken a specific custody protocol. Automated tooling can help defenders and attackers inspect code, spot unusual behavior, or test assumptions. The real risk depends on what attackers can access, what they can do, and how fast an organization can detect and patch defects.
The practical response isn't to assume every complex system will fail. It's to cut unnecessary complexity, require independent review, and keep a process for vulnerability reporting and updates. Offline signing can reduce exposure to some remote attack paths, but solid engineering and operational controls are still needed to handle flaws that exist inside the system itself.
What institutions should evaluate before adopting cold custody
A custody decision should start with the assets and the workflow, not with a product category. Holding Bitcoin as a long-term reserve is different from running an exchange, payments business, or fund that needs frequent transfers. Before comparing vendors, map how often assets move, who can request and approve transfers, how quickly transactions must be processed, and what recovery time is acceptable.
Then test whether the custody design fits that operating model. Ask the provider to explain where transaction data is created, how it reaches signing hardware, what operations run inside the HSM, and what stays outside it. Ask how approval authority is split and whether a single administrator can change policy or redirect requests. These questions turn broad claims like "secure" or "offline" into requirements you can actually test.
A practical review checklist
- Threat model: Which remote, insider, physical, and supply-chain attacks does the design address? Which risks remain outside its scope?
- Signing boundary: What exactly runs inside the HSM, and how can the institution verify that boundary in its own deployment?
- Authorization: How many approvals are required, who controls them, and can one person alter both policy and transaction data?
- Transaction verification: How do signers confirm the destination and amount using a trusted process before authorizing a transfer?
- Recovery: What happens after hardware loss, device failure, staff turnover, or an unavailable signer? Can recovery be tested without weakening normal controls?
- Operations: How long does a typical transfer take, and what happens during an urgent transaction or a network outage?
- Assurance: What independent reviews, audit evidence, and incident-response commitments can the provider share?
Match controls to asset use
A simple way to think about this is to separate assets by purpose. Long-term holdings can often tolerate stricter offline procedures and slower approvals because transaction frequency is low. Working balances usually need faster access - but that should come with explicit limits and monitoring. Make that separation real through policy and workflow, not just an informal distinction between "different wallets."
This approach accepts a trade-off: faster access typically requires more operational availability, while tighter isolation can add delay and complexity to normal transfers. The goal isn't to eliminate that tension. It's to decide where the organization can tolerate delays and where it can't, then set controls accordingly.
Run failure exercises before moving material funds
A written procedure isn't enough. Test routine transfers, rejected transactions, failed HSMs, an unavailable approver, and a recovery scenario. Check whether staff can follow the process without improvising, whether audit records are complete, and whether the organization can tell who authorized each action.
These tests often uncover problems that a purely cryptographic review misses: unclear ownership, a missing backup role, or a workflow that only works when a particular employee is available. Those are custody risks, too. Good custody tooling should make a sound operating model easier to enforce - not push the institution into undocumented workarounds.
Conclusion
Blockstream and Thales are making a specific architectural case: combine MPC-based authorization with dedicated hardware, and keep the signing protocol inside the HSM boundary. Adam Back's argument for offline custody is grounded in a real operational concern - connected systems create paths attackers may exploit, including through compromised employee devices and credentials.
But offline doesn't mean risk-free, and a simpler design isn't automatically a better deployment. Institutions should validate the signing boundary, test recovery, inspect approval controls, and measure the operational cost of isolation. For assets that rarely move, cold custody can be a sensible default. For active balances, the right setup may be a tightly limited, faster workflow alongside stricter long-term storage.
In the end, the decision should follow the asset's purpose and the organization's threat model. Map how funds move, decide where delays are acceptable, and test the full process before placing material Bitcoin holdings into custody.