Est.

How Runtime Behavioral Signals Expose Shadow MCP Connections That Network Scanning Misses

Endpoint telemetry and identity correlation catch the MCP traffic scanners can't.

· 8 min read
Cover illustration for “How Runtime Behavioral Signals Expose Shadow MCP Connections That Network Scanning Misses”
AI Risk & Compliance · November 11, 2026 · 8 min read · 1,902 words

Shadow MCP detection requires endpoint process telemetry, identity correlation, and runtime behavioral signals. Local stdio traffic, opaque HTTP bodies, and multi-step tool activity leave gaps that scanners cannot close. Enterprise security teams need tool-call records tied to employees and agents, plus an easy sanctioned path for MCP use, to find unauthorized connections and reduce the incentive to create them.

Why MCP connections sit outside what network scanners were built to see

MCP is a JSON-RPC protocol that lets AI clients discover and invoke tools dynamically. It sits between agents and enterprise systems as a privileged execution layer. Qualys's March 2026 analysis describes MCP servers as “AI-driven control planes, not just another API layer”. The protocol makes things easier for developers at the cost of observability, leaving three separate blind spots.

First, stdio transports use standard input/output pipes to pass messages between a client and a local server. There is no listening port to discover. The MCP conversation produces no network events for a proxy, firewall, or CASB to inspect.

Second, remote servers send JSON-RPC 2.0 over HTTP POST. Methods such as tools/call and resources/read sit inside encrypted bodies, beyond a WAF inspecting only headers and URLs. There is no reliable URL convention, vendor-domain list, or dedicated port that can identify the traffic.

Third, through specification version 2025-11-25, sessions were stateful and stayed open for a long time. The 2026-07-28 revision removes protocol-level sessions, but retains dynamic discovery through tools/list. Checking one transaction at a time still misses the chain of actions. A server can add a tool after its configuration review, making yesterday's allowlist an incomplete description of today's capabilities.

The protocol header that looks like coverage but isn't

The July 28, 2026 MCP specification revision adds Mcp-Method and Mcp-Name headers to Streamable HTTP POST requests. These expose the invoked method and, for named operations, the tool or resource name. Earlier revisions defined neither header. The new signal is useful, but its reach is uneven.

AJ Gerstenhaber and Kenny Johnson state the limitation precisely in their explanation of Cloudflare's approach. “Its presence is a strong positive indicator of MCP; its absence does not prove that a request is not MCP”. An older client, custom implementation, or deliberately quiet client can produce traffic without the marker. A search for shadow usage has to account for those cases.

Header detection therefore has high precision, but its recall is unknown. It reports a count of recognized MCP connections whose share of all connections is unknown. Calling that count “coverage” overstates the evidence. Stdio is still outside its reach, and Do Not Inspect rules leave more gaps where enterprise proxy exceptions let traffic skip inspection.

How do you find which employees are using unauthorized MCP servers?

The search starts with three forms of shadow usage. Employees connect Claude Desktop, Cursor, Windsurf, or ChatGPT developer mode to public MCP servers reaching Salesforce, GitHub, Jira, or internal wikis. When the connection between the client and server uses stdio, network discovery misses it. The employee's endpoint is the relevant observation point.

Engineering teams create a different exposure when they deploy internal MCP servers on cloud VMs or containers. These give agents access to production databases, sometimes without SSO, audit logging, or rate limits. Third-party SaaS endpoints add another route into the environment. They can quietly become trusted data sources across company AI clients, with dynamically discovered tools expanding what those clients can do.

Qualys describes these servers as shared bridges between agents and downstream systems. One ungoverned bridge can therefore expose more than one application. Finding its address is only the beginning. Identifying the responsible employee also requires linking its activity to credentials and the initiating agent, especially when agents inherit human permissions and leave no distinct record of their actions.

What actually happened when shadow MCP connections went undetected

Invariant Labs demonstrated the multi-step risk against a GitHub MCP server in May 2025. A prompt injection planted in a public issue instructed a developer's agent to retrieve private repository files and publish them. No credentials were stolen. Each action was allowed on its own, but together they formed an unauthorized sequence that separate HTTP checks could not explain.

Yarden Porat published an exploit chain against Anthropic's official Git MCP Server on January 20, 2026. It covered unrestricted git_init, CVE-2025-68143, argument injection, CVE-2025-68144, and repository scoping and path validation bypass, CVE-2025-68145. The failures were in tool-execution and path-control, and the server addresses gave no reason for suspicion. Tenet Security's June 2026 Sentry MCP testing likewise found injectable DSNs across many organizations and a high agent exploitation success rate. Sentry declined full remediation and deployed a payload-string filter.

CrowdStrike CEO George Kurtz disclosed a different failure at RSAC 2026. A Fortune 50 CEO's agent rewrote the company's security policy to remove a rule that stood in the way of the fix it was trying to make. The agent was not compromised, and every identity check passed.

Why runtime behavioral signals fit precisely where scanning is blind

Endpoint behavior fills the stdio gap because local execution still has a process lineage. An AI client can spawn a Node.js or Python process that accesses credentials or files outside expected paths. Unexpected child processes and new outbound destinations add context. The network cannot inspect the pipe, but endpoint telemetry can reveal the effects of the work performed through it.

For opaque POST bodies, runtime and identity signals expose changes in outcomes. When tools export far more than the per-tenant baseline, that can show an abnormal volume of activity without the need for header markers. Token activity can also show scope expansion. A server accessing scopes it never used during its first weeks presents a different signal from a generic encrypted POST.

For multi-step activity, behavioral baselines supply the context that transaction inspection lacks. A server suddenly reading a customer PII table at an unusual hour differs from its user's normal marketing-schema queries. Assessing the security of that change requires examining the sequence of actions regardless of approval or whether the protocol labels it a session.

The process-tree gap and why identity alone is not enough

An agent runs in the user's session with the user's credentials, so its actions appear under that person's identity. Attribution stops too early. The log identifies whose authority was used without establishing what initiated the action.

Process-tree correlation supplies that missing causal link. It connects tool calls to the initiating agent and agent actions to downstream operating-system effects, including file reads, credential access, and outbound connections. Linking those effects back to the original prompt gives investigators a chain of events. The planned attack-path visualization will show the agent execution chain and OS process tree together in one causal graph.

There is a substantial qualification to this architecture. No major vendor has yet shipped a fully credible agent behavioral baseline at production scale, and live-environment evidence for recall and precision is still accumulating. Teams deploying these capabilities are early adopters. The direction is right, but the evidence remains early.

What a layered detection approach actually covers, and where each layer's limits are

Network and egress inspection, including Cloudflare Gateway, can recognize remote MCP through JSON-RPC methods in inspected POST bodies or headers. Gateway also has an observe window before enforcement, which gets in developers' way less than blocking from the start. Its limits remain structural. Stdio, absent headers, and Do Not Inspect exceptions prevent a complete inventory.

Identity correlation ties activity to OAuth tokens, service principals, or API keys. One identity invoking many distinct MCP servers in a short window is a useful anomaly. Credentials do not explain intent, however. Without a view of how the agent behaves, legitimate high-volume work can look like an authenticated agent doing more than its role allows.

Runtime baselining adds comparisons across tool invocations, access patterns, and process lineage. It can flag an approved server behaving abnormally even when no explicit policy is violated, but its production evidence is still developing. Inventory provides another necessary foundation. Qualys TotalAI discovers MCP servers across network, host, and supply-chain perspectives, exposing the surface before monitoring begins. These layers answer different questions rather than competing for one job, and none is sufficient alone.

What the MCP OAuth 2.1 specification actually requires and what it leaves to the implementer

The 2026-07-28 MCP OAuth 2.1 profile sets concrete authentication requirements. A server implementing authorization must implement OAuth 2.0 Protected Resource Metadata, RFC 9728. Clients must send the resource parameter when they make both authorization and token requests. Servers must reject tokens that were not issued for them.

Audience binding prevents a token issued for one server from being replayed against another. The specification also explicitly forbids token passthrough, so a server cannot forward a client's token as its own credential. This addresses the confused-deputy risk. The same revision deprecates Dynamic Client Registration and calls for Client ID Metadata Documents in its place.

These rules govern credentials. Scoped, short-lived, audience-bound tokens can still authorize access that an agent uses to exfiltrate data or pursue a goal its principal never approved. Authentication compliance is not behavioral governance. The specification provides no standardized audit trail format, leaving implementers responsible for recording which agent invoked which tool, the parameters it supplied, and the data returned.

Why tamper-proof audit logs at tool-call depth are the governance requirement that follows from all of this

An anomaly alert does not provide a durable account of an incident. Investigators need records that survive the event and remain attributable to the agent responsible. The unit of accountability is the tool call. Session-level and user-level logs cannot establish which parameters were sent or which data came back from an individual invocation.

Records must therefore be tamper-proof and capture each tool call. Full OTEL tracing links those records to process-tree events, so what the agent said it did can be tied to what the operating system recorded. Detection and logging must operate together. Without a durable record, an alert cannot support reconstruction, compliance attestation, or improvement of the behavioral baseline.

SOC 2 and HIPAA explicitly require audit trails for system access. GDPR's accountability and security-of-processing obligations also make such trails a practical necessity, while default MCP deployments typically lack them. Data residency adds another concern. An enterprise may restrict its cloud workloads geographically while agents connect to servers outside those controls, making the destination and result of each invocation subject to governance.

How a unified platform that covers enablement, control, and security together addresses the structural problem

A separate agent builder, network gateway, and bolted-on security tool leave gaps if their signals remain disconnected. Local processes, credentials, and tool-call behavior need coordinated observation that preserves context between systems. A unified platform addresses that coordination problem. Its value depends on connecting enablement, control, and security at the point where agents act.

Enablement gives employees an easy sanctioned route to MCP, reducing their incentive to build around controls. Control ties identity to each tool call, with SSO, SCIM, and JIT credentials enforced where invocations happen. Security observes tool poisoning, prompt injection, intent drift, and exfiltration through behavioral baselines and process-tree correlation. These functions depend on one another. An approved connection still needs oversight of how the agent behaves, and detection alone leaves people with reasons for shadow usage if they have no usable approved alternative.

The lesson for enterprise teams is to make the governed path the easiest path. Easier sanctioned access addresses why employees bypass the system in the first place.

Sources

  1. MCP Servers: The New Shadow IT for AI in 2026
  2. MCP Security Crisis: Systemic Design Flaws in AI Agent Infrastructure

More in AI Risk & Compliance