Skip to main content
Skip to content
Agentic Finance

A 2-of-3 Wallet Does Not Guarantee a Human Veto: Recovery Design for AI Agents

A multisig threshold can leave the intended human optional. Map signer coalitions, modules and recovery paths before giving an AI agent wallet authority.

Human authorisation keys beside a vault, illustrating multisig approval and recovery paths
Reading options: audio, text size & sections

Decentralised News · 2027 planning series · Reviewed 21 November 2026

What Matters

A 2-of-3 multisig does not automatically give a human a veto over an AI agent. It accepts any two qualifying owner signatures. If the other two owners can act without the intended reviewer, that reviewer is optional. Modules and recovery permissions can create additional authorization paths, so the owner threshold is only the first part of the review.

The practical question is: which exact combinations of people, credentials and contracts can move funds or change the rules? This article provides a way to answer that question before giving an agent execution authority. It separates ordinary spending, emergency revocation and recovery, because each can have a different decision rule.

The Signal

Safe’s 1.5.0 release introduced Module Guards. That creates a concrete version question for agent wallets: which execution path and which guard are actually configured? DN’s worksheet begins with owner coalitions and keeps module and recovery review separate.

DN Alpha Thesis: recovery is another control path

DN’s analysis is that wallet governance should be evaluated as a set of executable paths. Counting signatures is useful, but it does not show whether every path respects the intended human boundary. A wallet can require several owner signatures for ordinary transfers while allowing a module to spend within a separate mandate or a recoverer to replace the owners after a delay.

That flexibility can make automation practical. It also means the phrase “human in the loop” needs a precise definition. Does a named person approve every transfer? Must any one of several humans participate? Or have humans approved a bounded policy in advance? Those are three different operating arrangements. A team should choose one explicitly and describe it honestly to everyone whose capital is exposed.

Evidence boundary: sources were reviewed on 21 November 2026. Safe provides the concrete documentation example. The coalition calculations and operating procedures below are original DN design analysis. DN did not create a wallet, inspect a reader’s deployment, audit a contract or execute a recovery.

1. Start with the rule you actually want

Safe documents its threshold as the minimum number of owners required to confirm an owner transaction. It is a numerical quorum, and the configured owners can change it through the account’s authorized process. That establishes how the ordinary owner path works; it does not assign special status to an owner because a person calls that owner “the human reviewer.” Safe smart-account concepts.

Write the intended rule as a sentence that can be checked. “Two approvals” is incomplete. “At least one human approves each transfer above the operating allowance” specifies more. “The treasury manager must approve every change to the execution policy” identifies a particular person and action class. “Any two independent human owners may replace a lost signer” defines an availability requirement.

Intended ruleWhat must be establishedCommon misunderstanding
Any two owners approveTwo eligible signatures satisfy the owner pathAssuming one named owner must participate
At least one human approvesNo accepted owner quorum consists entirely of nonhuman-controlled signersCounting two agent credentials as independent oversight
A specific human approvesNo permitted execution path can omit that person’s required approvalTreating a standard 2-of-3 threshold as a personal veto
Agent acts inside a prior mandateScope, amount, expiry and revocation are enforced by the relevant mechanismDescribing prior policy approval as approval of every later transfer
Humans can recover after key lossA surviving authorized path can change ownership without depending on the lost credentialAssuming a backup contact has executable recovery authority

These distinctions should appear in the deployment brief and the incident runbook. They determine which failure cases matter. A system designed for unattended small payments should not be rejected merely because it lacks per-payment approval. It should be evaluated against the bounded mandate it actually claims.

2. Enumerate the winning coalitions

Consider an invented account with three owners: Helen, Agent A and Agent B. Its threshold is two. The minimum winning coalitions are Helen plus A, Helen plus B, and A plus B. The third coalition can act without Helen. Calling Helen the supervisor changes nothing in this simplified threshold model.

Now replace Agent B with another human, Tariq. With Helen, Tariq and Agent A, every two-owner coalition contains a human. Helen still has no individual veto: Tariq and the agent can form a quorum. The account meets the “at least one human” condition for the ordinary owner path, while failing the stronger “Helen must approve” condition.

Illustrative owner setThresholdAny human required on owner path?Specific human guaranteed?
Helen, Agent A, Agent B2 of 3No: A and B form a quorumNo
Helen, Tariq, Agent A2 of 3Yes, assuming the stated control labelsNo: either human may be omitted
Helen, Agent A2 of 2Yes on this owner pathYes on this owner path
Helen, Tariq, Agent A3 of 3Yes on this owner pathYes, with lower tolerance of unavailable signers

The general arithmetic is straightforward. With N owners, threshold T, and H owners genuinely controlled by the required human group, a human-free threshold coalition exists when N − H ≥ T. For one specifically required owner, an ordinary threshold rule permits omission whenever the remaining owners can meet the threshold.

This is a logical test, not a probability of compromise. It assumes owners are distinct, the account uses the stated threshold path, and the control labels are accurate. An owner that is itself a smart account introduces its own nested authorization rules. A service controlled by a person is not automatically equivalent to that person reviewing a proposed transfer.

3. Independence matters more than the number of labels

Two agent processes can look like two approvers while sharing one cloud account, secret store, deployment administrator or model context. In that case, a single compromised dependency may control both. The same problem can affect nominally separate human credentials held on one device or recovered through the same account.

DN recommends recording both the signer address and its control domain. A control domain is the person or system whose compromise, coercion or outage could remove that signer’s independence. This is an analytical label, not a claim that a particular custody setup has been penetrated.

Ask a concrete question for each pair: what event could make both approvals unreliable at once? A shared administrator, copied recovery phrase, common signing service or unattended approval script should be visible in the answer. Separate model prompts alone are weak evidence of independent authority if the runtime can retrieve both credentials.

The coalition worksheet deliberately does not turn these answers into a security score. A count of independent domains can help identify concentration, but it cannot measure implementation quality or predict loss. For the wider custody questions, use DN’s wallet authority and key-security analysis.

4. Review modules as separate entrances

Safe modules extend the account with their own transaction logic. An authorized module can validate a request under its rules and call the account’s module-execution route. Safe warns that modules are security-critical and that a malicious module can take over an account. The owner’s ordinary threshold therefore does not describe every enabled execution route. Safe module documentation.

For each module, write down the action it can perform, who can trigger it, what limits it checks, and who can change or disable it. Separate a payment allowance from a general execution permission. Also separate a permission to propose a transaction from permission to execute one.

Safe’s current spending-limit documentation makes that last distinction explicit: a proposer prepares transactions for signer review, while an allowance lets a spender transfer the configured token within a cap without gathering owner signatures each time. Its documented allowance is not a recipient allowlist, and its cap is enforced in token units rather than continuously repriced dollars. Safe spending limits and proposer role.

A team could reasonably use a human-controlled treasury with a limited agent operating allowance. The accurate claim would be that humans govern the mandate and larger treasury actions. It would be inaccurate to say a human signs every payment from that allowance. Review all token allowances separately, then inspect contract approvals that may survive changes to the agent’s own credential.

5. Check the deployed version and guard configuration

Do not use the blanket rule “modules always bypass guards.” Safe’s September 2026 announcement identifies version 1.5.0 as the current Safe{Wallet} contract version and describes Module Guard support. Existing accounts may still use earlier deployments; the website’s current product version does not prove the version of a particular account. Safe 1.5.0 announcement.

The technical reference provides a separate setModuleGuard function. It checks module-initiated transactions, and a zero address disables that guard. A version that supports a control is therefore not evidence that the control is configured, or that its logic requires the intended human approval. Module Guard reference.

Guards can add checks around execution, but Safe also warns that a broken guard can block transactions. More restrictive policy can improve containment while creating an availability problem. A recovery design should therefore explain how the team handles a guard that refuses legitimate emergency actions. Safe guard documentation.

Record the chain, account address, contract version, owner set, threshold, enabled modules, guard addresses and observation block together. The useful conclusion is bounded: these controls were observed in this configuration at this point. An upgrade or module change should trigger another review, rather than silently inheriting the previous result.

6. Treat recovery as powerful authority

Safe’s documented RecoveryHub custom mechanism allows a preconfigured recoverer to propose replacement owners and a new threshold without satisfying the original owner threshold. Execution is delayed, and existing signers can reject a proposal if they can meet their threshold. The documented workflow requires onchain proposal and execution transactions; after the delay, execution can be performed by any account. These are properties of that described mechanism, not a universal definition of social recovery. Safe RecoveryHub documentation.

The design implication is important: a recoverer may eventually control who controls the funds. It belongs in the authority map even when it never makes an ordinary payment. If a recoverer can replace the required human, the human requirement is conditional on the recovery policy and the team’s ability to detect and challenge misuse.

A delay is useful only when somebody can act during it. DN’s proposed runbook records who monitors recovery proposals, who can satisfy the cancellation rule, how gas is funded, how an unavailable approver is contacted, and what happens if the primary interface fails. Verify current notification behavior instead of assuming an alert will arrive.

Choose the delay from operational requirements. A short delay improves recovery speed but reduces response time. A long delay creates more opportunity to react but may leave legitimate operations unavailable. Neither choice can be judged from the number alone; it must be compared with the team’s actual coverage and the time needed to assemble a valid response.

7. Use a recovery drill before raising the operating balance

The following is a proposed test plan, not a report of tests DN executed. Use a controlled environment and an approved deployment plan. The objective is to establish whether the written authority model matches the behavior that the team can reproduce.

DrillQuestion to answerEvidence to retain
One ordinary signer unavailableCan the permitted remaining group operate, and should it?Owner set, threshold and accepted or rejected authorization
Human approval omittedDoes the intended control reject that action on the relevant path?Exact action class, route and observed result
Agent credential revokedWhich actions stop, and which independent approvals remain?Executed revocation and subsequent permitted-path inventory
Unexpected recovery proposalWho notices, and can they assemble cancellation authority?Detection time, response participants and cancellation record
Primary interface unavailableCan authorized owners inspect and respond through an approved fallback?Fallback procedure, dependencies and unresolved gaps
Configuration changedDid owners, modules or guards alter the previous conclusion?Before-and-after configuration and new coalition analysis

Failing a drill is useful evidence. Write the narrow failure: the monitoring path did not alert, the cancellation group could not form, or a surviving allowance remained active. “Recovery failed” is too vague to tell the next operator what to fix.

8. DN Human-Veto and Recovery Worksheet

The accompanying instrument asks for owner count, the threshold, which owners belong to the required human group, and whether the requirement applies to any human or a named person. It enumerates the minimum owner coalitions and reports whether one can omit the required approval.

A separate configuration section records modules and recovery. Unknown paths remain unresolved. A passing owner-coalition result cannot clear an unreviewed recovery mechanism or certify a guard. This separation avoids converting a small mathematical result into a broad security endorsement.

In the Helen/A/B example, the tool should show three minimum coalitions, including one without Helen. Changing the threshold to three removes that owner-path omission, but all three owners become necessary for ordinary execution. Adding a separately authorized module does not alter the coalition arithmetic; it changes the scope of the conclusion.

The tool accepts descriptive aliases rather than private keys or recovery phrases. Keep sensitive credential records in the team’s approved security system. A worksheet should explain authority without becoming another repository of secrets.

DN Human-Veto and Recovery Worksheet

Change the illustrative assumptions below. This worksheet runs locally in your browser. It records your declarations; it does not verify evidence, connect to accounts, submit instructions or grant approval.

Required approval can be omitted on the owner path

Owner-path threshold2 of 3
Owners labeled human1
Minimum owner coalitions3
Coalitions omitting required approval1
Maximum absent owners leaving some quorum1
  • This enumerates the ordinary owner threshold only. Labels, key independence, deployed contracts and configuration are not verified.
  • Module or recovery authority is unknown or unreviewed. The owner-path result cannot establish an overall human veto.
  • The absent-owner count states only that some ordinary quorum remains possible. It is not a guarantee for a particular outage, required human, common controller or recovery event.
Agent A + Agent B — required approval omitted Helen + Agent A — required approval included Helen + Agent B — required approval included

The initial numbers are synthetic. Unknown evidence remains unresolved. Downloads contain the values you enter; use aliases and avoid private or identifying information.

9. Decide where the agent belongs

For a workflow that only proposes transfers, an agent may not need owner authority at all. A proposal can include the destination, amount, reason, source evidence and expected state change for authorized signers to review. For repetitive low-value activity, a separately bounded execution policy may be more practical than asking people to approve every routine call.

For treasury governance, policy changes and recovery, keep the approval requirement explicit and independent of the model’s instructions. The system should not be described as having a human veto simply because a human configured it once. Its authority must be checked on the route used by the action being discussed.

Read this alongside DN’s ERC-4337 agent-account guide and spending-limit evidence worksheet. The former explains the account stack; the latter tests what a stated cap actually covers. This article adds the missing coalition and recovery boundary.

Methodology and limitations

DN reviewed the primary sources linked above and compared this question with existing wallet and account-abstraction coverage. Product statements are documented capabilities, not independently observed deployment results. All named people and signer sets in the worked examples are invented. The arithmetic enumerates ordinary threshold coalitions; it does not model cryptographic exploits, malicious signing interfaces, coercion or nested contract-owner behavior.

Labels such as “human-controlled” are reader declarations. They cannot establish that a person understood a transaction or that two credentials are operationally independent. Module, guard and recovery review requires the exact contracts and configuration. No live account has been connected by the article, and no funds have been moved.

Sources and evidence register

Primary sources reviewed on 21 November 2026. The article distinguishes documented facts, DN analysis and synthetic examples. Recheck changeable rules and configurations before acting.

  1. Safe smart-account concepts — Ordinary owner threshold, owners and signatures; reviewed 21 November 2026.
  2. Safe modules — Separate module execution and security-critical authority.
  3. Safe guards — Additional execution checks and denial-of-service risk.
  4. Safe RecoveryHub custom recovery — Recovery changes owners/threshold after delay; existing threshold can cancel. Article includes legacy release-era wording, so do not claim current UI/notification behavior without deployment verification.
  5. Safe 1.5.0 release announcement — September 22, 2026 announcement; module guards are version/configuration specific.
  6. setModuleGuard reference — Separate module guard; zero address disables it.
  7. Safe allowances and proposer role — Proposer cannot move assets alone; allowance cap token-denominated, not recipient allowlist.

Frequently asked questions

Does a 2-of-3 multisig guarantee human approval?

No. Any two eligible owners can satisfy the ordinary threshold. With one human and two agent-controlled owners, the two agents can form a quorum without the human.

What is the difference between any-human approval and a named human veto?

Any-human approval requires at least one member of a human group in every accepted coalition. A named human veto requires that particular approval; a standard threshold may allow other owners to omit that person.

Can a recovery mechanism change a human-veto rule?

Potentially. A recovery path that changes owners or threshold can change the future approval rule. Its delay, cancellation authority and monitoring must be reviewed separately.

Do Safe modules always bypass guards?

No blanket statement is accurate. Safe 1.5.0 supports Module Guards. The account's deployed version, enabled guard and actual policy must be checked; support for a control does not prove it is configured.

Does a passing worksheet mean the wallet is safe?

No. It establishes only a logical result for the declared owner threshold and labels. It does not audit contracts, verify identities or clear unreviewed module and recovery paths.

DN research note: This article is original editorial analysis supported by the linked sources. Calculations and example cases are illustrative unless explicitly described otherwise. The worksheet is a planning aid; it does not carry out the actions it describes.

Get the most talked about stories directly in your inbox

Join the Decentralised News briefing for independent crypto, DeFi and AI analysis. No spam, unsubscribe anytime.