Asset Mgmt MCP Challenges | Moring
The main MCP governance challenges for asset managers are shadow servers connected without review, credentials that outgrow their purpose, governance frameworks that do not cover runtime tool calls, and content-injection and supply-chain attacks on the tool layer. A structural protocol flaw disclosed in April 2026 left roughly two hundred thousand instances potentially vulnerable, underscoring that the tool layer needs its own controls rather than inherited ones.
Short answer
The main MCP governance challenges for asset managers are shadow servers connected without review, credentials that outgrow their purpose, governance frameworks that do not cover runtime tool calls, and content-injection and supply-chain attacks on the tool layer. A structural protocol flaw disclosed in April 2026 left roughly two hundred thousand instances potentially vulnerable, underscoring that the tool layer needs its own controls rather than inherited ones.
How it works
The first challenge for asset managers is that MCP arrived without permission. Most enterprises did not decide to adopt the Model Context Protocol; engineers did, one configuration file at a time, and the result is a class of asset that reads internal wikis, opens pull requests, queries production databases, and posts to collaboration channels at machine speed. In investment firms the same pattern reaches order management, research repositories, and client records. Security teams typically learn about server connections after the fact, which inverts the approval process and leaves the firm governing an estate it did not choose.
The second challenge is credential sprawl. Every MCP server connection creates a new credential, such as an API key, OAuth token, database connection string, or service account secret, and most enterprises have no standardized framework for credential type, lifetime, scope, provisioning, rotation, revocation, or exposure detection. Surveys report that fifty-three percent of MCP servers use static API keys, the weakest available pattern. For an asset manager, each ungoverned credential is a standing path into systems carrying fiduciary and client obligations, and revocation maps rarely match the reality of what is deployed.
The third challenge is that existing governance does not reach the tool layer. Most enterprise AI governance frameworks, including the EU AI Act, were built to govern models: training data, outputs, risk classification, transparency, and accountability. None were designed to answer for an agent's runtime decision to call an MCP tool, so the tool layer is effectively ungoverned even in organizations with mature AI programs. A firm can have its EU AI Act risk tiers mapped and its model review board operating under SR 11-7 and SR 26-2 influenced practice and still be unable to say which of its MCP servers were ever vetted for anything.
The fourth challenge is a new threat model. Research on securing MCP identifies three adversary types that exploit its flexibility: content-injection attackers embedding malicious instructions into otherwise legitimate data, supply-chain attackers distributing compromised servers, and agents that become unintentional adversaries by over-stepping their roles. Because MCP increases attack surface through dynamic, user-driven integrations, threats arrive through the data the agent reads, not only through the network perimeter, which defeats controls that inspect packets rather than semantic intent.
Protocol-level fragility compounds all of it. In April 2026, a disclosure documented remote code execution against six live production platforms through a structural flaw in MCP's design, a flaw Anthropic formally declined to fix at the protocol level, leaving roughly two hundred thousand potentially vulnerable instances. More than eighty percent of Fortune 500 companies run MCP in active production workflows, so the flaw is not hypothetical for any large institution, asset managers included. The implication is architectural: assume the protocol is imperfect and put compensating controls at the gateway.
Finally, governance platforms that were never built for agents let new tool surfaces appear without review; analyses of enterprise platforms note that MCP servers can expose new capabilities to agents without tenant-level review, bypassing controls such as data loss prevention and environment segmentation. The remedial path is unglamorous: inventory what exists, classify it, score risk, and enforce policy at a control plane the agents cannot modify.
Frequently asked questions
Why is the shadow MCP problem harder to fix after adoption?
Once engineers have connected servers one configuration file at a time, the firm is retrofitting governance onto an estate it did not approve, often without knowing its true size. Connections reach production databases, code repositories, and collaboration tools, and revoking them can break workflows people now depend on. Asset managers should start with discovery and inventory, classify each server by the data and systems it touches, and then move approved servers behind a gateway while sunsetting the rest, rather than attempting a big-bang cutoff that breaks live research or operations work.
What makes the MCP threat model different from ordinary API security?
MCP exposes three adversary types that static integrations do not: content-injection attackers embedding malicious instructions in legitimate data, supply-chain attackers distributing compromised servers, and agents that over-step their roles and become unintentional adversaries. Because tool calls carry semantic intent, network controls that inspect packets cannot judge them. For an asset manager, an agent reading market commentary could ingest an injected instruction to exfiltrate holdings data, so defenses must validate tool responses, isolate execution, and require human approval for consequential actions rather than relying on perimeter security alone.
How serious is the disclosed protocol-level vulnerability for asset managers?
A disclosure in April 2026 documented remote code execution against six live production platforms through a structural flaw in MCP's design, which Anthropic formally declined to fix at the protocol level, leaving roughly two hundred thousand instances potentially vulnerable. With over eighty percent of Fortune 500 companies running MCP in production workflows, exposure is likely in any large institution. Asset managers should treat the protocol as imperfect by design and add compensating controls at the gateway, including interception, response scanning, and rate limiting, rather than waiting for protocol-level remediation that is not coming.
Why do existing AI governance programs fail to govern the tool layer?
Programs built on frameworks such as the EU AI Act govern models: training data, outputs, risk classification, and transparency. They were not designed to answer for an agent's runtime decision to call a tool, so the tool layer remains ungoverned even where a firm has EU AI Act risk tiers mapped and a model review board meeting monthly under SR 11-7 and SR 26-2 influenced practice. Closing the gap requires extending inventory, classification, and risk scoring to MCP servers and grants, treating each server as a governed asset with its own approval and evidence record.