Est.

How to Tell Whether a Vendor Actually Understands MCP Threat Anatomy or Just Relabeled Their CASB

Six MCP attack classes give buyers questions that expose a relabeled CASB.

· 8 min read
Cover illustration for “How to Tell Whether a Vendor Actually Understands MCP Threat Anatomy or Just Relabeled Their CASB”
AI Risk & Compliance · November 10, 2026 · 8 min read · 1,858 words

An enterprise MCP security and governance platform must control each tool call, track agent identity through delegation, detect runtime manipulation, and record why each action was allowed. This article explains how to distinguish those capabilities from a relabeled CASB. The deciding evidence is how a vendor answers questions about MCP’s attack mechanisms.

Why MCP threats cannot be detected by pattern-matching controls

A vendor that leads with CASB integration as an MCP answer is describing a control that cannot see the attack. MCP merges reasoning and control flows within a shared semantic context, where metadata, returned data, and executable instructions coexist without strong isolation. An attacker can exploit what the model understands. That places the control surface in the shared semantic context.

A poisoned tool call can produce valid HTTPS to permitted endpoints, using legitimate authentication tokens and syntactically correct JSON-RPC. The request looks normal because it is normal. Its danger comes from the instruction that caused the model to make it, which need not produce a known-bad signature or anomalous file movement. A WAF or DLP rule looking for those signals has no basis for recognizing the semantic manipulation.

Identity and visibility gaps deepen that mismatch. The MCP authorization specification defines an OAuth 2.1 framework but explicitly makes authorization optional, leaving room for servers without authentication. CASB access governance, meanwhile, is built around human identity rather than autonomous agents that use inherited credentials across many services. Session identity cannot explain every delegated tool call. Traditional DLP and CASB/SASE controls also miss local stdio workflows, agents embedded in development environments, and API-based AI tool usage, leaving those deployments outside their view.

The specific attack classes that define the MCP threat surface

Vulcan Research / AIFT’s March 2026 MCP-38 taxonomy maps 38 threat categories to STRIDE, OWASP Top 10 for LLM Applications, and OWASP Top 10 for Agentic Applications. It covers tool manifests, input schema fields, stdio transport, and multi-server routing that prior frameworks miss. Six attack classes give buyers a practical vocabulary for testing that understanding.

Tool poisoning embeds malicious instructions in apparently normal tool metadata or responses, placing them in the model’s context as trusted input. NYIT’s March 2026 threat model identifies it as the most prevalent and impactful client-side MCP vulnerability. Five of seven clients tested had no static validation against it. Users also typically do not see complete tool descriptions at runtime, even when approval dialogs technically expose them. Can the vendor detect instructions that the model receives but the user overlooks?

Rug pulls change a tool’s description or behavior after approval, when tool lists are fetched again at runtime. UltraViolet Cyber’s TIDE and AST teams documented AIRedScam, in which a threat actor forked AIScan-N and embedded a SmartLoader infostealer payload inside a compressed archive, with the payload arriving post-install. The vendor’s question is therefore about timing. What detects a change after the clean tool has already been approved?

Tool shadowing takes advantage of the context window that servers share. A server without protected-data access can steer the agent toward another server that has it, even though it has no credential for that server. Confused deputy attacks cross a related permissions boundary when a server exercises its own broad privileges on the requesting user’s behalf. Invariant Labs demonstrated a crafted GitHub issue that hijacked an assistant into leaking private repository data through a public pull request. Can the platform distinguish permitted access from manipulated use of that access?

Multi-agent delegation and permission escalation expose another boundary. Standard bearer token formats do not prevent a sub-agent from reusing a delegated credential at full scope, although each hop should only reduce permissions. Supply chain mutation brings changes from outside the workflow itself. NUS researchers’ September 2025 taxonomy describes malicious servers as cheap to generate and difficult for existing scanners to detect. The corresponding questions are whether delegation narrows authority and whether monitoring continues after installation.

What genuine MCP security architecture has to do that CASB-era thinking never required

Tool-level authorization must evaluate each invocation against what this agent may do in this context. On-Behalf-Of (OBO) authentication lets the agent exercise the exact permissions of the human who initiated the request. Agent identity must also persist through the workflow so that the platform can attribute individual calls.

Delegation requires a separate constraint on credential scope. The MCP security specification warns that broad-scope access tokens can leak through logs, memory scraping, or local interception. Even without theft, delegated credentials create an authorization problem if a sub-agent can reuse their full scope. Bearer tokens cannot enforce narrowing on their own. Other controls must keep each delegation hop from gaining more authority than it received.

Continuous tool-definition integrity monitoring checks for changes between when a tool is approved and when it runs. Hash-based monitoring detects changed definitions, including rug pulls and description changes associated with tool shadowing, even when the altered tool is never invoked. Runtime inspection serves a different purpose. It examines tool responses as they enter the agent’s context, while behavioral monitoring looks for intent drift between the user’s goal and the agent’s subsequent actions. Definition integrity does not replace semantic detection.

The MCP OAuth 2.1 profile has no standardized audit trail format, so vendors must supply records connecting agents, users, tools, parameters, and policy decisions. Much of this infrastructure lacks finalized standards. MCP-I / KYA-OS is in active community development with sparse production implementations, while the CSA AI Trust Framework is a governance framework. “Compliance-ready” claims therefore often promise architecture against unfinished standards, a gap the diagnostic questions must expose.

The governance and compliance pressure that makes vendor selection consequential now

The EU AI Act’s high-risk obligations apply at the tool-calling layer to MCP gateways that are used on regulated data. A subsequent regulation that entered into force in 2026 deferred the enforcement deadline. That deferral does not provide the missing controls. To meet those obligations, a retrofitted CASB must be able to link tool calls to users and record policy decisions.

The operational pressure is already visible in deployment outcomes. Only 11–14% of enterprise agentic AI pilots reach production, with the rest stalling on identity, audit, and access-control gaps. Agent capability is not the bottleneck described here.

Cloud Security Alliance’s May 2026 research note says to treat every MCP server as an untrusted third party, apply zero trust at the tool integration layer, and put ongoing governance in place [1][2][3]. UltraViolet Cyber’s May 2026 advisory calls for version pinning, signed tool definitions, continuous dependency monitoring, and quarterly red-team exercises targeting agentic workflows. Multi-agent workflows also require permissions to account for both the delegating and receiving agents without escalating trust, which human-identity CASB architecture does not provide.

How do you evaluate an enterprise MCP security and governance platform?

A genuine answer identifies where the control operates. A retrofitted answer substitutes an existing security function without explaining how it reaches the MCP-specific failure.

How does the platform detect a rug pull, and what happens at detection? A genuine answer describes hash-based or cryptographic fingerprinting of tool definitions at the proxy layer, with detection, logging, and blocking before the changed definition reaches the agent. Timing is part of the answer. Installation scans, server allowlists, and network monitoring do not detect runtime description mutation. The follow-up is whether monitoring covers changed descriptions on servers the agent never directly invokes but that share its context.

How is an agent process identified through every tool call? A genuine answer explains how non-human identities are created, scoped, periodically reviewed, and decommissioned, as well as how identity propagates through a multi-step workflow. Authentication at session start is only one step. A service-account mapping or human SSO identity at session level does not attribute individual calls to the initiating user across delegation. The follow-up is whether Agent B receives explicitly constrained permissions when Agent A delegates to it.

Can the platform allow tool X but deny tool Z on the same server, based on the initiating human’s identity? A genuine answer describes OBO authentication at the tool-call level, with the agent exercising the requesting user’s permissions. A retrofitted answer describes server or endpoint allowlisting, which admits or denies the entire server.

How does the platform prevent a sub-agent from reusing a delegated credential at full scope? A genuine answer explains what bearer tokens cannot enforce on their own and names the compensating controls that narrow permissions at every hop. Credential protection is a different control problem. Vaulting, secrets management, and token rotation address credential security at rest or in transit.

Does the audit log record the policy decision permitting each call? A genuine answer describes structured records of which agent acted for which user, under which policy, invoking which tool with which parameters. Session recording and SIEM integration are not substitutes for that policy state. The follow-up is whether logs are tamper-proof and include full OTEL tracing so a call can be reconstructed across the complete workflow.

How does the platform find prompt injection inside a tool response? A genuine answer describes runtime inspection as responses enter the context window, plus behavioral anomaly detection for actions that diverge from the initiating user’s goal. Input validation, regex filtering, and DLP scanning do not show that the platform can spot instructions made to look like normal response data. The follow-up is how it detects a server steering the use of another server with protected-data access.

What happens when a clean installation’s upstream repository or package registry entry is later compromised? A genuine answer describes continuous dependency monitoring, proxy-enforced version pinning, and signed tool definitions. Approval cannot freeze the upstream supply chain. Pre-deployment scans, vendor approval workflows, and allowlisted registries address the earlier onboarding decision. AIRedScam gives this question a concrete test case because its payload arrived after installation.

What to do when a vendor's answers reveal a gap — and what a sound evaluation process looks like from here

When answers leave things out, buyers should check what the platform enforces today and ask what controls it has beyond authentication. Auth0’s “Auth for MCP,” generally available May 6, 2026, and Okta’s MCP server address the authentication layer. Authentication is necessary, but it is not governance. Neither layer substitutes for tool-level authorization, runtime semantic detection, or complete audit records. Likewise, an MCP-native gateway can control reachable tools without inspecting the model reasoning that drives selection, so runtime behavioral monitoring must complement gateway controls.

Whatever stage the vendor has reached, security and platform teams should list the MCP servers across their environment and note which deployments are local and which are remote. They should also turn off tool-call auto-approval in production and require OAuth 2.1 with least-privilege scoping on every server connection. These steps establish a concrete governance baseline. They follow the advisory and CSA guidance rather than depending on a vendor’s future standards support.

The strongest evaluation outcome is a vendor explaining its architecture precisely and acknowledging uncovered areas. Product boundaries also deserve scrutiny because delegation and audit attribution can break across seams between an agent builder, gateway, and separate security tool. A unified platform covering enablement, control, and runtime security reduces those integration gaps. Teams that make these distinctions rigorously become the teams that enable AI adoption at scale.

Sources

  1. Model Context Protocol Threat Modeling and Analyzing Vulnerabilities to Prompt Injection with Tool Poisoning
  2. When MCP Servers Attack: Taxonomy, Feasibility, and Mitigation
  3. MCP-38: A Comprehensive Threat Taxonomy for Model Context Protocol Systems (v1.0)
  4. MCP Security Crisis: Systemic Design Flaws in AI Agent Infrastructure

More in AI Risk & Compliance