Asset Management Agent Identity Conflict Resolution | Moring
Conflicts in agent identity governance arise when permissions, delegations, or behaviors contradict each other: an agent holds access its purpose no longer supports, two agents' authorities overlap on a sensitive system, or an agent acts within its grants against the intent behind them. Resolution requires an owner-level decision recorded against evidence, with separation-of-duties checks and escalation when owners disagree.
Short answer
Conflicts in agent identity governance arise when permissions, delegations, or behaviors contradict each other: an agent holds access its purpose no longer supports, two agents' authorities overlap on a sensitive system, or an agent acts within its grants against the intent behind them. Resolution requires an owner-level decision recorded against evidence, with separation-of-duties checks and escalation when owners disagree.
How it works
Identity conflict in agentic environments is subtler than conflicting human access requests, because agents can disagree with their own paperwork. Identity governance analysis describes the core pattern: autonomous identities can chain decisions, delegate across systems, and stay inside existing permissions while still acting beyond what anyone intended. For an asset manager, this means a conflict may not look like two clashing entitlements; it may look like one agent, correctly provisioned, whose observed behavior contradicts the purpose recorded at registration. Resolution must therefore begin with detection, not just adjudication.
Detection comes from the intent-to-execution loop. Identity governance research describes behavioral verification as continuously reconciling what an agent is authorized to do with what it actually does, and every mismatch it surfaces is a candidate conflict: a research agent reading order-system data, a reporting agent exporting datasets it was never meant to move, an automation whose usage pattern no longer matches its registered purpose. Each detected conflict should arrive with the identity, the owner, the grants in force, and the observed actions, so resolution starts from a complete picture rather than an alert alone.
The first resolution path is purpose realignment. Because agents persist after the business context that justified their privileges changes, identity framework research warns, many conflicts are stale authorization rather than malice: the mandate closed, the project pivoted, the vendor changed, and the permission set stayed. The owner decides whether to rescope the identity to its current purpose, retire it, or formalize the expanded purpose with a new classification. Whichever choice is made, the decision and rationale are recorded against the identity, so the next reviewer sees a resolved conflict, not a rumor.
The second path is separation-of-duties arbitration. When two agents' authorities overlap on a sensitive system, for instance one agent that prepares valuation inputs and another that validates them, the conflict is structural, and grant-level fixes cannot resolve it. The firm should apply the same separation logic it uses for human roles: no identity, human or agent, should both prepare and approve the same consequential artifact, and delegation chains must not let an agent effectively approve its own work. Resolution means redesigning the permission sets, recording the segregation decision, and adding the pair to a conflict matrix so future requests are checked against it automatically.
The third path handles intent conflicts that grants cannot express. Security research demonstrated how an agent using a legitimate authenticated session was hijacked by hidden instructions in a webpage it summarized, acting against its user's interests while remaining technically authorized. This class of conflict resolves through input-side controls rather than permission edits: sanitizing tool responses, inspecting content before it reaches the model, and tightening what external sources an identity may consume. Owners should treat each such incident as a configuration defect with a defined fix, not a one-off anomaly.
When owners disagree, or the conflicted system is too consequential for the owning team to decide alone, escalation decides. Runtime governance research recommends automatic escalation paths for refusals and anomalies, and identity conflicts touching order flow, valuations, or client communications warrant routing to a control committee with the evidence attached. Escalation outcomes should feed back into policy: a recurring conflict pattern signals that permission packaging or classification needs adjustment, so the resolution process improves the system rather than merely closing tickets.
The compliance thread runs through documentation. Model risk disciplines under SR 11-7 and continued under SR 26-2 expect the firm to demonstrate accountable, independently reviewable control over systems influencing decisions, and conflict resolution records, the mismatch, the analysis, the decision, the rescoping, are exactly that demonstration for agents. Where European operations or vendors bring the EU AI Act into scope, documented intervention in automated behavior supports deployer oversight duties. An asset manager that treats identity conflicts as recorded decisions rather than informal fixes turns a messy operational reality into examinable governance.
Frequently asked questions
What kinds of identity conflicts arise with AI agents at an asset manager?
Three main kinds. Stale authorization conflicts, where agents persist after the business context that justified their access has changed, as identity framework research warns. Structural conflicts, where two agents' authorities overlap on a sensitive system in ways that undermine separation of duties. And intent conflicts, where an agent stays inside its grants while acting beyond anyone's intended purpose, the pattern identity governance analysis describes as chaining decisions within existing permissions. Each type requires different evidence and a different resolution path.
How should a detected conflict between an agent's grants and its behavior be resolved?
Start with purpose realignment backed by the intent-to-execution evidence. The owner reviews what the agent is authorized to do, what it actually did, and what its current business purpose is, then either rescopes the identity to match the purpose, formalizes the expanded purpose with a new classification, or retires the identity. Whichever decision is made, record it with its rationale against the identity. Detection alone is not resolution; the recorded decision is what makes the conflict auditable and prevents silent drift from recurring.
How should overlapping agent authorities on a sensitive system be arbitrated?
With separation-of-duties logic applied to agents as rigorously as to people. No identity, human or agent, should both prepare and approve the same consequential artifact, and delegation chains must not allow an agent to effectively approve its own work. When overlap is found, redesign the permission sets, record the segregation decision, and add the conflicting pair to a conflict matrix checked automatically on future requests. For overlaps touching order flow, valuations, or client communications, route the decision to a control committee rather than leaving it with one owning team.
What should be done when an agent is hijacked despite correct permissions?
Treat it as an input-side configuration defect. Security research demonstrated agents being steered by hidden instructions planted in content they legitimately processed, acting under valid authenticated sessions, so permission edits alone do not resolve the class. Remediate by sanitizing tool responses, inspecting content before it reaches the model, and narrowing which external sources the identity may consume, then record the incident and fix as a versioned configuration change. Owners should review whether similar content channels feed other agents, since one successful hijack pattern rarely stays unique.