Asset Mgmt MCP Compliance | Moring
MCP compliance for asset managers means demonstrating that agent tool activity is inventoried, approved, scoped, and evidenced under the firm's model risk and AI oversight obligations, including SR 11-7 and SR 26-2 influenced practice and the EU AI Act. Because no major framework covers runtime tool calls, compliance rests on the firm's own records: server onboarding, per-tool grants, gateway logs, and approvals.
Short answer
MCP compliance for asset managers means demonstrating that agent tool activity is inventoried, approved, scoped, and evidenced under the firm's model risk and AI oversight obligations, including SR 11-7 and SR 26-2 influenced practice and the EU AI Act. Because no major framework covers runtime tool calls, compliance rests on the firm's own records: server onboarding, per-tool grants, gateway logs, and approvals.
How it works
Compliance for MCP in an asset manager begins with an uncomfortable fact: the frameworks the firm is accountable to were not written for this layer. Most enterprise AI governance frameworks, including the EU AI Act, were built to govern models, covering training data, outputs, risk classification, transparency, and accountability, and none were designed to answer for an agent's runtime decision to call an MCP tool. The tool layer is therefore ungoverned by default even in firms with mature programs, and compliance work must map agent tool activity onto existing obligations rather than wait for new rules.
The mapping is straightforward in structure. Model risk governance under SR 11-7 and SR 26-2 influenced practice expects material systems to have identified risk, documented controls, and accountable owners; an MCP server that reaches order management, research, or client records is such a system. Server onboarding decisions, tool permission grants, and gateway policies should be documented with the same rigor as model documentation, and the EU AI Act's expectations for oversight and logging of systems that influence decisions attach to the same records. The compliance artifact set is: inventory, classification, approval, scoping, enforcement, and logs.
Evidence generation is where most programs fall short, and regulated-enterprise guidance is blunt about why: MCP servers enter environments through developer toolchains, not procurement, creating a structural gap between deployments and governance documentation. Closing it means the enforcement layer produces evidence as a side effect. Vendors in this market emphasize exactly that, evidencing compliance against the frameworks a firm is accountable to while auto-discovering, registering, and inventorying servers and blocking non-compliant ones from production. Compliance teams should demand that capability rather than assembling evidence manually after deployments occur.
The threat landscape shapes the compliance argument too. Research on securing MCP identifies content-injection attackers embedding malicious instructions into legitimate data, supply-chain attackers distributing compromised servers, and agents that over-step their roles as the adversary types exploiting MCP's flexibility. An examiner asking how the firm addresses these will accept a coherent answer: response scanning, third-party server validation, least-privilege tool grants, and human approval for consequential actions, each mapped to a named control. The absence of that mapping is what turns a technical gap into a compliance finding.
A disclosure in April 2026 documented remote code execution against six live production platforms through a structural flaw in MCP's design that Anthropic declined to fix at the protocol level, leaving roughly two hundred thousand instances potentially vulnerable. For compliance purposes this is a supervisory expectation in waiting: firms running the protocol, as most large institutions do, should be able to show compensating controls at the gateway and an assessment of exposure, in the same way they document compensating controls for any third-party dependency with a known defect.
Finally, compliance is a cadence, not a document. Server inventories drift, grants accumulate, and protocol versions change; quarterly review of the catalog, drift alerts, and post-incident replays keep the record current. The aim for an asset manager is that any examination question about agent tool activity can be answered from the system of record within a day, which is the practical definition of MCP compliance.
Frequently asked questions
How do SR 11-7 and SR 26-2 influenced obligations apply to MCP servers?
Model risk governance expects material systems to carry identified risk, documented controls, and accountable owners, and an MCP server that reaches order management, research, or client records qualifies. Asset managers should document server onboarding decisions, tool permission grants, and gateway policies with the same rigor as model documentation, and assign each server an owner. That mapping turns unfamiliar technology into familiar evidence, letting the firm extend existing oversight instincts to the tool layer without inventing a parallel governance regime for agents.
What evidence should an asset manager keep for MCP compliance?
Keep six artifacts: the server inventory with owners, the classification of each server by data sensitivity, the approval record for each server reaching production, the per-agent and per-tool grant matrix, the gateway logs of every call with arguments, results, and identities, and the records of human approvals for consequential actions. Retention should match the firm's examination standards. Because frameworks including the EU AI Act were not designed for runtime tool calls, this firm-generated evidence set is what demonstrates oversight; without it, the tool layer has no compliance story at all.
Why is the deployment-versus-documentation gap a compliance risk?
MCP servers enter through developer toolchains rather than procurement, so deployments outrun governance documentation structurally, not occasionally. An examiner or internal audit will read that gap as a control environment that permits change without approval, regardless of how good the firm's model governance looks on paper. Asset managers close it by making the enforcement layer generate evidence automatically, through inventory, contract generation, and call-time logging, so documentation cannot lag deployment, and by blocking unapproved servers from production rather than documenting them after the fact.
How should the disclosed protocol vulnerability factor into compliance?
The April 2026 disclosure documented remote code execution against six production platforms through a structural flaw Anthropic declined to fix at the protocol level, leaving roughly two hundred thousand instances potentially vulnerable. Compliance treatment should mirror third-party risk practice: assess the firm's exposure, document compensating controls at the gateway such as interception, response scanning, and rate limiting, and record the decision. Supervisors increasingly expect exactly that pattern for dependencies with known defects, and an agent tool layer is now a dependency firms must be ready to defend.