Maker-Checker Separation of Duties & RFC 8693 Delegation for High-Risk Agent Actions
Why 'human-in-the-loop' breaks when the agent owner clicks Approve—and how RFC 8693 actor claims and WebAuthn step-up authentication close the loop.
Key Takeaways
- Agent owners and the initiator of a delegated RFC 8693 token chain are automatically blocked from approving their own agent's mutations.
- Approver and TenantAdmin decisions enforce step-up MFA via OIDC acr and amr claims before votes are accepted.
- Per-agent deactivation cancels Pending/Approved actions and blocks new evaluations and redeems; already executing writes cannot be recalled. Compromised votes can be revoked without erasing audit history.
Stopping self-approval across delegation chains
A common gap in approval portals is checking only the immediate caller identity. If Alice delegates authority to an agent using OAuth 2.0 Token Exchange (RFC 8693), the token carries an act claim linking the agent back to Alice.
The Gateway resolves the initiating human subject from the verified delegation token and rejects malformed or mismatched delegation. The Control Plane blocks the recorded requester and agent owner from approving; this is not an arbitrary delegation-chain authorization engine.
Step-up MFA at the moment of decision
Holding a long-lived browser session is not strong enough to authorize a €250,000 invoice or a supplier bank-account change. Policy enforcement checks OIDC Authentication Context Class Reference (acr) and Authentication Methods References (amr) claims at decision time.
Outside Development, the Control Plane requires an accepted acr or amr MFA claim and returns 403 when it is absent. The default accepts OTP as well as WebAuthn; it does not enforce authentication recency or automatically initiate a browser step-up. Configure and test those requirements at the identity provider.
- Two-person dual control configurable per tool and amount threshold
- Per-agent deactivation cancels Pending/Approved actions; there is no tenant-wide kill-switch endpoint
- Compromised-approver revocation invalidates tainted votes and records an immutable revocation event
A worked example: who may approve Alice's agent?
Consider a hypothetical finance team. Alice owns an AP agent and, from her own session, delegates it a task to change the bank account of supplier SUP-1001 from a DE89 IBAN to an SE45 IBAN. The token the agent presents keeps Alice as the subject and names the agent as the actor in the act claim. The policy says two approvals are required.
When votes arrive, the control plane builds the exclusion set for this action: the requester, the registered agent owner, and every principal in the act chain. Alice is in that set twice over, as owner and delegator, so her vote is refused. Bob and Carla, both approvers outside the chain, each complete a step-up challenge. Their votes bind to the action digest. If someone edits the IBAN after Bob votes, the digest changes and his vote no longer counts.
Edge cases that break separation of duties in practice
Most SoD failures are organizational, not cryptographic. The rules above only help if the approver pool reflects real independence.
- Shared or functional accounts: an approver identity used by several people defeats the point. Require personal OIDC identities.
- Small teams: with N-of-M set to 2 and only two eligible approvers, one absence stalls every action. Size M with holidays in mind.
- Role changes: when the agent owner changes, the exclusion set should follow the current owner and the owner recorded at request time.
- Approver also builds the agent: treat the engineer who deploys an agent as an owner for SoD purposes.
What the kill switch does and does not halt
A kill switch blocks new submissions and execution claims. An action already redeemed and executing at that moment completes or ends as UNKNOWN, and is then reconciled rather than interrupted mid-write.
When dual control is the wrong tool
Two-person approval costs reviewer time on every action. Reserve it for changes that are hard to undo, such as supplier bank details or large invoice approvals, and let deterministic Allow rules handle routine, bounded, low-value actions within per-agent quotas.
Security scope and deployment prerequisites
This is a pilot implementation, not a zero-risk guarantee or compliance certification. Examples describe repository capabilities or deployment recommendations, not proof that a live installation has enabled them.
Every enrolled write fails closed during control-plane outages. Privileged control-plane MFA is mandatory outside Development, but accepted MFA claims do not prove recent authentication or hardware-backed WebAuthn unless the identity provider enforces those properties.
Per-agent deactivation blocks new evaluations and redeems and cancels Pending/Approved actions; it cannot recall an already executing ERP write. Evidence verifies recorded events, not unrecorded ERP activity; independently retained heads and trusted keys are needed to detect truncation or replacement.
Production blockers remain
The static portal still uses sessionStorage tokens (BFF not implemented). Off-host backup signing, immutable backup storage, live IdP MFA/logout testing, restore drills, adapter validation, monitoring, and legal review require deployment-specific work. Anchoring and SIEM forwarding are optional and must be configured and tested. See the repository deployment guide and runbook gap register before using real customer data.
Related deep dives
Defending Against AI-Assisted Invoice Redirection Fraud: Two-Person IBAN & BIC Verification
Nov 21, 2026 · 3 min readSecurity & AuditGRC Automation (Vanta, Drata) vs. Cryptographic Decision Evidence Vaults
Nov 20, 2026 · 3 min readArchitectureiPaaS Connectors (Workato, Zapier, Make) vs. Fail-Closed Agent Gateways
Nov 19, 2026 · 3 min read