Asset Mgmt Agent Governance Risks | Moring
Agent risk in asset management concentrates in delegated authority: over-permissioned credentials, injected instructions, data-layer blind spots, and multi-step actions that compound. Sweet catalogs failure surfaces including data access, prompts, models, tools, identities, and APIs; Atlan cites Gartner's finding that traditional governance misses 60 to 70 percent of agent-specific risk vectors. Naming the risks precisely is the precondition for governing them.
Short answer
Agent risk in asset management concentrates in delegated authority: over-permissioned credentials, injected instructions, data-layer blind spots, and multi-step actions that compound. Sweet catalogs failure surfaces including data access, prompts, models, tools, identities, and APIs; Atlan cites Gartner's finding that traditional governance misses 60 to 70 percent of agent-specific risk vectors. Naming the risks precisely is the precondition for governing them.
How it works
The risk inventory for agents begins with what makes them categorically different from prior AI. Palo Alto Networks describes the structural shift: earlier systems generated insights while human operators decided what to do next, whereas agents increasingly execute tasks directly within business workflows, which is why governance must establish explicit oversight and accountability for agent behavior. Akamai adds the compounding mechanism: agents plan multi-step tasks, invoke tools, and act without continuous human oversight, introducing compounding risks that traditional AI governance was never designed to address. For an asset manager, each of these properties lands on fiduciary processes that were engineered assuming a human closed every loop.
Sweet's guidance provides the most concrete failure taxonomy: agent failures involve injected instructions, leakage, and over-permissioned credentials, and controls must therefore span data access, prompts, model choices, tool calls, identities, and APIs. Each item maps to a recognizable asset management scenario: a prompt-injected research agent exfiltrating position data, a tool call reaching an order system beyond the agent's mandate, a shared service account erasing attribution, an API integration carrying more privilege than the workflow requires. Sweet's runtime test, expected, permitted, and reversible, is designed against exactly this list rather than against abstract model quality.
The largest blind spot sits below the behavior layer. Atlan's two-layer framework argues governance must cover both what agents can do and what data they can consume, and that most enterprise frameworks govern agent behavior while leaving the data layer ungoverned, where most AI failures originate; it cites Gartner's finding that traditional AI governance misses 60 to 70 percent of agent-specific risk vectors, especially at data and context. Bigeye's analysis makes the complementary point that authorization controls and access logs leave unhandled whether the data the agent acted on was trustworthy. For asset managers, stale research inputs or mis-scoped client data flowing into an otherwise well-permissioned agent is precisely the unexamined path to a consequential error.
Organizational risk completes the picture. BCG's CIO guide observes that agents are proliferating under inconsistent rules and fragmented oversight as business units adopt agents from multiple vendors and employees build personal agents, raising one of the most urgent governance challenges of the moment; Rakesh Gohel's analysis adds the ownership pathology, asking who owns agent risk among security, governance, and compliance and concluding that if the answer is all three, it is probably no one, supported by findings that 92 percent of organizations lacked proper AI access controls. Bitsight's framing supplies the diagnostic questions: which AI systems are in use, what data can they access, who owns them, and which create the most risk.
The mitigation posture follows the inventory rather than chasing headlines. Microsoft's Cloud Adoption Framework baseline closes identity and access exposure before agents operate; Sweet's per-action runtime tests bound delegated authority continuously; Atlan's data layer brings input quality and classification under governance; discovery per Zenity's checklist eliminates the shadow-agent population where several of these risks incubate. Anchor the register in SR 26-2 and SR 11-7 conventions for investment-relevant consequence and the EU AI Act's requirements for high-risk systems, and the risk register stops being a document and becomes the index of a working control set.
Frequently asked questions
What makes agent risk categorically different from earlier AI risk?
Delegated execution. Palo Alto Networks describes the shift from systems that generated insights while humans decided next steps to agents executing tasks directly within business workflows, and Akamai adds that agents act without continuous human oversight across multi-step tasks, compounding risk in ways traditional governance never addressed. For asset management this lands on fiduciary processes engineered assuming a human closed every loop: an agent can complete a research-to-execution chain without any human between steps. Risk registers built around model output quality therefore miss the operative danger, which is authorized-seeming action.
Which concrete failure surfaces must agent controls cover?
Sweet's guidance catalogs them: data access, prompts, model choices, tool calls, identities, and APIs, with failures involving injected instructions, leakage, and over-permissioned credentials. Each maps to an asset management scenario: a prompt-injected research agent exfiltrating position data, a tool call exceeding its mandate on an order system, a shared service account erasing attribution, an over-privileged integration. Sweet's expected, permitted, and reversible test is built against this list, which is why it functions as the risk register's operational spine rather than an abstract control slogan.
Where do traditional governance programs miss agent risk?
In the data layer. Atlan's two-layer framework holds that governance must cover both what agents do and what data they consume, since most frameworks govern only behavior while the data layer stays ungoverned, where most AI failures originate, citing Gartner's estimate that traditional AI governance misses 60 to 70 percent of agent-specific risk vectors, especially at data and context. Bigeye adds that authorization controls leave unhandled whether the data acted on was trustworthy. For asset managers, stale or misclassified inputs flowing into a well-permissioned agent is the archetype of this blind spot.
What organizational conditions amplify agent risk?
Fragmented ownership and ungoverned proliferation. BCG's CIO guide describes agents spreading under inconsistent rules and fragmented oversight as business units adopt multiple vendors and employees build personal agents, and Rakesh Gohel's analysis concludes that when security, governance, and compliance all own agent risk, no one does, citing 92 percent of organizations lacking proper AI access controls. Bitsight's diagnostic questions, which systems are in use, what data they can access, who owns them, which create the most risk, expose the gaps. Microsoft's baseline and Zenity's discovery method are the structural remedies. Central baseline and discovery close the organizational gaps asset management faces.