Model Context Protocol standardizes how AI agents receive tools, resources, and prompts from external servers through a client-server architecture. Tool descriptions, contextual data, and templated instructions are all passed directly into a model's context window at runtime. The MCP specification does not mandate that this content be treated as untrusted; it relies on the host application to build its own trust and safety boundaries.
This distinction matters because prompt injection in a single-model LLM application typically involves one input channel: the user prompt, possibly combined with retrieved documents. MCP introduces multiple independent sources of untrusted content in a single session, including tool metadata from one or more servers, dynamic resource payloads, and server responses that may be updated after initial integration. Each of these can carry natural-language text that a model interprets as an instruction rather than as data, simply because MCP does not require labeling which parts of context came from tools versus users versus system instructions.
How the Attack Works Against MCP Specifically
Industry security research from 2024 and 2025 consistently describes three primary injection vectors specific to MCP. The first is malicious tool descriptions, sometimes called tool poisoning, where a server exposes a tool whose description field contains hidden instructions designed to manipulate the connected agent. Since tool descriptions are ingested as plain text into the model's context, there is no structural distinction between a legitimate capability description and an embedded command.
The second vector is poisoned resource content. When an agent calls a tool that returns a file, an API result, or other data, that response can contain natural-language instructions the model may follow as if they were directives from the user or system. The third vector involves compromised or impersonated MCP servers. Because MCP does not require cryptographic signing or provenance verification of tool descriptions or server responses in its base specification, an agent has no built-in way to confirm that a server's content has not changed since it was last trusted.
A documented pattern in MCP security research is the rug-pull scenario, where a server's tool descriptions are approved by a user or client at one point and then altered afterward, changing behavior once trust has already been established. Static review at integration time does not catch this kind of later-introduced change.
Why the Impact Extends Beyond Text Generation
Prompt injection against a traditional single-model LLM application generally results in manipulated or leaked text output. MCP changes the exploitability and blast radius of a successful injection because the protocol is designed to let agents invoke tools, not just generate responses. If an attacker can get malicious instructions into an agent's context through a tool description or resource response, the practical result can be an unauthorized API call, unintended file access, or data transmitted to an external destination.
This expands the consequence of injection from an output-quality problem to an action-authorization problem. Combining multiple MCP servers in a single agent session compounds the risk further, since each additional server is another untrusted content source merged into the same context window, and any one of them can potentially influence tool calls directed at the others.
Governance Considerations for Enterprise Deployments
No regulatory framework currently names MCP specifically, but existing AI governance obligations apply to it in general terms. NIST's AI Risk Management Framework Generative AI Profile calls for documented risk controls around third-party model and tool integrations, which is directly applicable to how enterprises vet MCP servers. OWASP's Top 10 for LLM Applications identifies prompt injection as its top risk category and recommends treating all external tool and plugin outputs as untrusted input requiring validation before use in downstream decisions.
Because MCP does not natively verify the provenance of tool descriptions or server responses, enterprises need to assign clear accountability for vetting third-party servers and for continuously monitoring them after approval. This is a governance gap the protocol leaves open by design, not an oversight that will necessarily be closed by future specification revisions.