Insecure by Default: Study Finds 1 in 5 MCP Access Policies Broken or Missing

New research reveals systemic authorization failures in the Model Context Protocol ecosystem as maintainers rush to patch fundamental security gaps.

edit
By LineZotpaper
Published
Read Time2 min
A recent study has found that roughly one in five access policies associated with the Model Context Protocol (MCP) are either broken or missing, raising serious concerns about the security of a protocol increasingly used to connect AI assistants to enterprise tools and data.

The findings come from researchers at an unnamed firm who analyzed MCP integrations across customer and prospect environments over the past several months. They discovered that over 20 percent of MCP-related access policies they reviewed were non-functional or absent, leaving the door open for unauthorized access to internal systems.

This vulnerability is compounded by the findings of a 2026 benchmark called MCPTox. Tested against 45 live MCP servers and 20 models, it revealed an average attack success rate of 36.5 percent, rising to 72.8 percent against the worst-performing model. The benchmark measured how easily an AI agent could be tricked by malicious instructions hidden in tool descriptions — a vector demonstrated as early as May 2025 by researchers at Invariant Labs, who showed that GitHub's own MCP server could be hijacked through a poisoned public issue.

The systemic weakness is further illustrated by ecosystem-wide data. Across the MCP ecosystem, 88 percent of servers require credentials to function, but only 8.5 percent actually use OAuth. By some estimates, over half of MCP servers in the wild run on static API keys or personal access tokens that are rarely rotated, and close to half of enterprise AI activity runs through personal accounts rather than service accounts.

A concrete example of this problem surfaced with Splunk's own MCP Server app, which logged session and auth tokens in cleartext until patched in version 1.0.3 (tracked as CVE–2026-20205).

On July 28, 2026, the Model Context Protocol's maintainers shipped a specification update built almost entirely around authorization: issuer validation, issuer-bound client credentials, and Client ID Metadata Documents. The update is widely seen as an admission that the original trust model did not survive contact with production environments.

A common failure mode identified by the research is that scopes default to everything. Many servers that need only read-only calendar access request read, write, and admin permissions because that is what tutorials use. In practice, most MCP tools are never scoped in the first place, often reusing admin-level API keys originally created for other purposes years prior.

§

Analysis

Why This Matters

  • The MCP protocol is becoming a standard for connecting AI assistants to enterprise tools and data. These flaws mean that even simple — and increasingly common — internal tools built by non-experts can expose sensitive systems to attack.
  • The gap between the number of tools that need credentials (88 percent) and those using robust OAuth (8.5 percent) suggests the ecosystem is operating on an honor system that has demonstrably failed.
  • The prevalence of static, non-rotating tokens and personal accounts handling enterprise AI activity means a single compromised credential could give an attacker persistent access to an entire organization's AI-managed infrastructure.

Background

The Model Context Protocol (MCP) is an open protocol designed to standardize how AI assistants securely access data and tools. It aims to simplify integrations, allowing non-expert developers to build powerful AI-driven workflows rapidly — as illustrated by the common example of an employee wiring together a scheduling tool with internal APIs in an afternoon. The protocol's maintainers ship updates, and a growing ecosystem of servers supports integration with everything from Slack and GitHub to CRMs. The security overhaul published on July 28, 2026 — introducing issuer validation and client registration standards — marks a direct response to production failures.

Key Perspectives

MCP Maintainers: The specification update released July 28, 2026, indicates their awareness that the original trust model needed a fundamental redesign with issuer validation and bound credentials. Researchers (Invariant Labs, MCPTox team): Their work demonstrates that tool descriptions are an active attack surface — unvalidated text can hijack an agent's behavior, and testing shows high success rates against even modern models. Critics/Skeptics: The fundamental architecture trusts the AI agent to interpret tool descriptions safely, which may be an unsolvable problem that no specification update can fully fix.

What to Watch

  • Adoption rate of the new authorization specification among MCP server developers
  • Whether the MCPTox benchmark's attack success rates decline with model updates or new safeguards
  • Emergence of tool-safety linters or sanitization standards for MCP tool descriptions

Sources

newspaper

Zotpaper

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.