AI Agent Security8 min read

MCP Security Risks Every Enterprise Should Know

The Model Context Protocol is connecting AI agents to enterprise systems at scale. But MCP servers introduce new attack surfaces — from credential exposure to prompt injection relay. Here's what security teams need to address.

TigerIdentity Team|

The Model Context Protocol (MCP) has gone from niche experiment to enterprise standard in under a year. Originally introduced by Anthropic to give AI assistants structured access to external tools and data, MCP is now the de facto protocol for connecting LLMs to databases, APIs, file systems, and cloud services.

The productivity gains are undeniable. Developers use MCP servers to give Claude or GPT access to their codebase, documentation, and deployment pipelines. Sales teams connect agents to CRM data. Finance teams build MCP integrations for real-time reporting. The ecosystem is booming.

But enterprises are adopting MCP faster than they're securing it. Most MCP servers run with no authentication, store credentials in plaintext configuration files, and grant agents unrestricted access to the systems they connect to. For security teams, this is a ticking time bomb.

The MCP Explosion

10K+
MCP Servers Published
78%
Fortune 500 Experimenting
3x
MCP CVEs in Q1 2026

5 MCP Security Risks You Can't Ignore

1

Credential Exposure in MCP Server Configs

MCP servers require credentials to access the systems they connect to — database passwords, API keys, OAuth tokens. Most store these in plaintext JSON configuration files on developer laptops or CI/CD runners.

Scenario: A developer's laptop is compromised. The attacker finds MCP config files with production database credentials, Slack bot tokens, and AWS access keys — all in plaintext.
2

Overly Permissive Tool Definitions

MCP tools are defined with broad capabilities by default. A "database" MCP server typically exposes SELECT, INSERT, UPDATE, and DELETE operations — even when the agent only needs to read data. Tool definitions rarely scope access to specific tables or columns.

Scenario: An analytics agent connected via MCP has write access to the production users table. A prompt injection causes it to execute a bulk UPDATE that corrupts customer records.
3

Prompt Injection Relay

MCP servers return data from external systems to the LLM. If that data contains malicious instructions — embedded in emails, documents, or database records — the LLM may follow them. MCP creates a bridge between untrusted data and privileged tool execution.

Scenario: An attacker places prompt injection text in a support ticket. When an agent reads the ticket via MCP, the injected instructions cause it to exfiltrate customer data through another MCP tool.
4

No Authentication Standard

The MCP specification does not mandate authentication between clients and servers. Most community MCP servers run on localhost with no auth. In enterprise deployments with remote MCP servers, there's no standard way to verify who is connecting or what permissions they should have.

Scenario: An internal MCP server providing access to HR data runs without authentication. Any process on the network can connect and query employee records, salary data, and performance reviews.
5

Supply Chain Risks

The MCP ecosystem is largely community-driven. Thousands of open-source MCP servers are published on npm and GitHub with minimal vetting. Enterprises install these servers to connect agents to Slack, Jira, Confluence, and other tools — granting third-party code access to production systems.

Scenario: A popular MCP server for Salesforce integration contains a backdoor that exfiltrates OAuth tokens to an external endpoint. It's discovered 4 months after installation.

How TigerIdentity Secures MCP

TigerIdentity's MCP Gateway sits between AI agents and MCP servers, adding authentication, authorization, and audit to every tool invocation — without modifying your existing MCP servers.

MCP Gateway Proxy

All MCP traffic routes through TigerIdentity's gateway. Agents connect to the gateway instead of directly to MCP servers. The gateway handles authentication, evaluates policies, and proxies authorized requests to the target server.

Tool-Level Authorization

Every MCP tool invocation is evaluated against policy in real-time. Policies can restrict which tools an agent can call, what parameters it can pass, and what data it can access — all based on the agent's identity and context.

Credential Vaulting

MCP server credentials are stored in TigerIdentity's encrypted vault — not in plaintext config files. Credentials are injected at runtime with automatic rotation. Agents never see the actual credentials.

Behavioral Analysis

ML models track normal tool usage patterns for each agent. Unusual patterns — sudden spikes in database queries, access to new tables, or tool calls outside typical workflows — trigger automatic throttling or session revocation.

MCP Access Policy Example

Here's how TigerIdentity governs MCP tool access. This policy allows an analytics agent to read data via MCP but blocks writes and restricts access to sensitive tables:

policy "mcp-analytics-agent":
  description: "MCP tool access for analytics agents"

  principals:
    type: ai_agent
    attributes:
      agent_type: analytics
      mcp_enabled: true

  default_decision: deny

  rules:
    # Allow read-only database queries
    - name: database-read
      effect: allow
      resources:
        type: mcp_tool
        server: database-server
        tool_name: query
      conditions:
        - mcp.params.query starts_with "SELECT"
        - mcp.params.table not in ["users_pii", "payments", "credentials"]
        - time.is_business_hours() == true
      rate_limit:
        max_requests: 100
        window: 1h

    # Block all write operations
    - name: block-writes
      effect: deny
      priority: 100
      resources:
        type: mcp_tool
        server: database-server
      conditions:
        - mcp.params.query starts_with ["INSERT", "UPDATE", "DELETE", "DROP"]

    # Allow Slack read, block Slack write
    - name: slack-read
      effect: allow
      resources:
        type: mcp_tool
        server: slack-server
        tool_name: [search_messages, get_channel_history]

    - name: slack-write-block
      effect: deny
      resources:
        type: mcp_tool
        server: slack-server
        tool_name: [send_message, upload_file]

Secure Your MCP Infrastructure

Add authentication, authorization, and audit to every MCP tool invocation. TigerIdentity's MCP Gateway deploys in minutes with zero changes to your existing servers.

Free for up to 10 agents. No credit card required.