Your agent is wired to a row of MCP servers. When it needs a dataset it picks a tool named data_fetch_and_store: right name, right description, compatible arguments. But which server is it actually calling? A team at the Canadian Institute for Cybersecurity (CIC), University of New Brunswick, posted a measured study on arXiv on November 16 (2611.21577): five frontier models acting as MCP clients, and three of them picked the attacker-controlled fake tool in 50 out of 50 trials.

The attacker only edits its own advertisements

The failure mode is called tool-provider misbinding: a request intended for a legitimate provider gets resolved to an attacker-controlled tool. The attacker compromises nothing — not the host, not the model, not the network channel. It simply runs one more MCP server and writes whatever metadata it likes: the tool name, the natural-language description, the input schema. The evaluation is a full factorial over three factors — name (unrelated or exact), description (unrelated, copied, or authority-enhanced with wording like enterprise-approved and officially verified), schema (incompatible or compatible) — twelve conditions in total. Ten prompts, five repetitions each, 50 trials per condition per model, and no trial failed to execute. The single legitimate tool, data_fetch_and_store, does exactly what it says: fetch a dataset and store it in the destination. Note that half of the ten prompts deliberately used security-aware wording that stressed trusted sources — as we will see, that did not save them.

Name, description, and schema all aligned: near wipeout

The worst cluster is C10 — exact name, copied description, compatible schema. GPT-5.4, Qwen3.7 Max, and DeepSeek V3 all hit a violation rate of 1.00, and Gemini 2.5 Pro reaches 0.84. A rate of 1.00 means all fifty trials went to the attacker, including the five security-aware prompts. More informative is C09, which differs from C10 only in schema: the average violation rate is 0 with an incompatible schema and about 0.78 once it becomes compatible. The paper calls schema compatibility the strongest common amplifier. Exact names alone do nothing: C07, an exact name with an unrelated description and an incompatible schema, sits essentially at zero across all five models. And a name collision is not even required: C06 uses an unrelated name with an authority-enhanced description and a compatible schema, and DeepSeek V3 violates at 0.50, Gemini 2.5 Pro at 0.42. The attack surface is wider than duplicate names.

Five models, five reactions

GPT-5.4, Qwen3.7 Max, and DeepSeek V3 rarely ask; they just pick. The paper describes DeepSeek V3 as the profile with the highest mean violation rate and almost no clarification. Claude Sonnet 5 takes another route: in C10 its violation rate is only 0.06, but clarification hits 0.94 and resolved-binding stability is 0.00 — not wrong, but not bound either; the ambiguity is pushed back to the user. Gemini 2.5 Pro mixes both behaviors, and the paper calls its combined profile the weakest in this experiment. The pattern survives perturbation: across five unrelated names and five independently written authority-enhanced descriptions, Qwen3.7 Max and DeepSeek V3 stay at 1.000 under C12. Raising the attacker budget from one candidate to ten leaves C10 and C12 saturated — a ceiling effect, as the authors note — but pushes GPT-5.4 in C07 from 0.04 to 0.28 and Qwen3.7 Max in C08 from 0.40 all the way to 1.00.

The root cause sits in the protocol

The MCP specification itself is explicit: tools are advertised by name, description, and inputSchema, and clients must consider tool annotations untrusted unless they come from trusted servers. The protocol already concedes that this metadata is untrusted; what it lacks, as the paper reads it, is a mandatory mechanism that binds a tool to a verified provider identity. The closest analog, the authors argue, is namespace confusion in software supply chains — package systems learned long ago that a familiar name is not a publisher. The proposed fix is correspondingly plain: provider-bound tool identity, concretely signed tool manifests, stable provider-bound identifiers, schema hashes, and client-side enforcement.

Cold water

The authors list their own limits: one controlled dataset-fetching task, five clients, and no claim to have exhausted attacker strategies — this affects how far the rates generalize, not the core finding that provider-unbound metadata produces measurable wrong-provider execution. The v1 preprint ships no code repository. And the same group previously published a journal-scale threat model of four agent protocols including MCP; this paper turns one of those risks into a measurable experiment, not an independent replication.

As agents connect to more and more tools, which tool to call is becoming a real security boundary. Next time your agent picks one in a blink, ask a simple question: what makes it so sure about that particular shop?