- Unit 42 tested an AI agent with five MCP servers connected and recorded a 78.3% attack success rate, VentureBeat reports. The reporting does not define a successful attack.
- The reported mechanism is cascading trust: instructions planted by one compromised server can reach later calls to servers that were never compromised.
- Treat each MCP server as an OAuth resource server, use short-lived narrow tokens, and pin tool definitions so changes are caught.
Trust that travels with the conversation
An AI agent that uses tools acts on what it reads. That is the weakness at the heart of this story. The Model Context Protocol (MCP) is a standard that links agents to business systems. The risk is not one flaw in one product. It is how trust moves between tools once they are connected.
What the testing found
Unit 42, the research team at Palo Alto Networks, tested an agent with five MCP servers connected. It recorded a 78.3% attack success rate. VentureBeat reported the result on November 15. It adds that five servers is ordinary for a team that has moved beyond a first pilot.
The reporting gives one figure, for five servers. It gives no figures for other server counts. It does not say what counts as a successful attack. It does not describe how Unit 42 ran the test. Read the number as a single test result, not a measure of live deployments.
How one server can mislead the rest
VentureBeat calls the mechanism "cascading trust". The agent treats what one server returns as reliable context. If that server is compromised, it can plant instructions in that context. The agent then carries those instructions into its later tool calls. Some of those calls go to servers that were never compromised.
Think of a forged note slipped into a shared case file. Nobody broke into anyone's desk. Every colleague who reads the file simply acts on the note. The agent is the colleague, and the tool output is the file.
A protocol that added security in stages
The MCP project's changelog for its November 2025 revision is one example of the rules tightening. It recommends a new way for clients to register, called Client ID Metadata Documents. It lets a client ask for extra permission in steps rather than all at once. It says servers must answer with HTTP 403 Forbidden when a request carries an invalid Origin header. It also updates the Security Best Practices guidance.
The same revision adds features. Sampling gains tool calling. Experimental support for tasks lets a client track long-running requests. New capability and new rules arrive together.
VentureBeat reads the changelogs as five stages, and its later stages go beyond what this one changelog lists. First, the spec had no authorization framework. Next came one based on OAuth 2.1. Then servers were classed as OAuth resource servers. Resource indicators (RFC 8707) became mandatory, to stop a malicious server from collecting tokens meant for other services. After that, Client ID Metadata Documents arrived as an option. In the latest stage, VentureBeat says, the spec deprecated dynamic client registration in favour of them, and clients had to check the iss parameter (RFC 9207).
VentureBeat's summary is that MCP launched with no authentication requirements, and each revision since has added them. It says servers built in the first year probably rest on assumptions the spec has since left behind.
The gap is outside the protocol
VentureBeat cites NSA guidance from May 2026, mapped against the OWASP MCP Top 10. By that mapping, only two of the ten risk categories sit in the content an agent reads before it acts. The other eight cover matters such as credentials, scope and identity. They sit in the surrounding infrastructure, where the protocol sets no built-in limits.
OWASP puts token mismanagement and secret exposure first in its priority order. VentureBeat's point is that MCP standardized how agents find and call tools. Proving who is calling, and limiting what they can reach, falls to the organisation running the agent. The buying question is not only which agent to deploy. It is who that agent may act as at each step, and who checks.
What Microsoft's guidance suggests
VentureBeat describes two Microsoft patterns. The first is for internal servers. Microsoft's own internal MCP governance hands out short-lived tokens with minimal rights, even for traffic that stays inside. The logic is that internal does not mean trusted.
The second is for servers open to customers, such as remote or multi-tenant services. Microsoft's Azure guidance suggests a gateway that checks identity. It turns away any call that lacks a valid token issued for that specific service.
The same guidance suggests a registry that pins each tool's definition. This guards against a "rug pull", where a trusted dependency quietly changes behaviour. Drift detection flags the change before production does. Microsoft also warns against granting "full mailbox access where read-only would do".
Questions to put to your team
Ask how many MCP servers each agent connects to, and who approved each one. Ask whether every server validates tokens as an OAuth resource server. Ask whether clients check the iss parameter. Ask how long tokens last and how narrow their scope is.
Ask whether tool definitions are pinned, and who is alerted when one changes. Ask which servers were built in the protocol's first year and have not been reviewed against the current spec.
On the mechanism Unit 42 tested, an agent can be only as trustworthy as the least trusted server it listens to.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Model Context Protocol project (modelcontextprotocol.io), the official MCP specification changelog.




