When tool names are not identity: measuring provider misbinding in MCP-based AI agents
When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents
The authors formalize MCP tool-provider misbinding as an attacker-controlled tool selection problem, and measured violation rates of 1.00 for GPT-5.4, Qwen3.7 Max and DeepSeek V3 under condition C10 across 5 LLM clients.
The attacker does not compromise the host, client, LLM, legitimate server or network channel; they merely operate another MCP server visible to the same client and advertise tools with self-chosen metadata (name, description, schema); the goal is for requests that should go to the legitimate provider to be resolved to the attacker-controlled tool (tool-provider misbinding violation).
- MCP lets clients combine tools advertised by multiple servers, but the tool metadata used for discovery is not necessarily bound to a verified provider identity, and may resolve requests intended for a legitimate provider to an attacker-controlled tool.
- The authors formalize tool-provider misbinding as an attacker-controlled tool selection optimization problem, conduct a 2×3×2 factorial experiment (12 conditions) on tool name, description, and schema in a controlled MCP environment, and run robustness experiments on different names, description wording, and attacker budget.
- Across 5 LLM clients, the combination of same-name conflict + task-aligned description + compatible schema (C10, C12) produces the highest violation rates; under C10, VR is 1.00 for GPT-5.4, Qwen3.7 Max, and DeepSeek V3; Claude Sonnet 5 shifts more to clarification requests; without same-name conflict, authority-enhanced description plus compatible schema (C06) can also cause non-zero violations.
- The authors note that they evaluated only one controlled data-fetching task and five clients, that the metadata variants were designed to isolate factors, and that they did not exhaust all attack strategies and real deployments; the paper does not report the clients' specific implementations, model versions or sampling parameters.
- Evaluated models
- GPT-5.4Claude Sonnet 5Gemini 2.5 ProQwen3.7 MaxDeepSeek V3
- Benchmarks
- 受控 MCP 环境(一个合法服务器 + 一个攻击者控制服务器,数据集获取任务,12 个攻击者工具条件 C01–C12)
- Metrics
- VR(violation rate)CR(clarification rate)FRLSRRBS(resolved binding stability)
The paper studies the tool provider misbinding problem in MCP agents: the metadata used to discover and select tools is not itself bound to a verified provider identity. An attacker does not need to compromise the host, the LLM, a legitimate server, or the network channel; they only need to run another MCP server that publishes tools with metadata of their choosing, and requests that should go to a legitimate provider can then resolve to an attacker-controlled tool. The authors formalize this failure as an attacker-controlled tool selection optimization problem and evaluate five LLM-based clients in a controlled MCP environment. The results show interaction-driven misbinding: failures are strongest when exact name conflicts, task-aligned descriptions, and compatible schemas reinforce one another; authority-enhanced descriptions that do not rely on name conflicts, together with compatible schemas, can also lead to execution by the wrong provider or to unresolvable provider ambiguity. Different models expose the same binding weakness in different ways, including attacker-controlled provider selection, requests for clarification, and unstable or inconsistent provider resolution.
Related events and news
1 events · 1 news mentions
News that mentions this paper
预印本称多个LLM客户端在MCP中误选仿冒工具
- TechTimescites the paper's id or link预印本:名称描述schema均匹配时,GPT-5.4等三模型50次全选仿冒MCP工具
GPT-5.4, Qwen3.7 Max and DeepSeek V3, according to a preprint posted November 16 on arXiv.
Paper analysis
What problem does this paper attempt to solve?
MCP tool metadata is not bound to a verified provider identity, and the authors study when this "tool-provider misbinding" leads to execution by the wrong provider.
Core problem: The metadata (name, description, schema) that MCP clients use to discover and select tools is not necessarily bound to a verified provider identity, and may resolve requests intended for a legitimate provider to an attacker-controlled tool.
- Scenario and importance: LLM applications are shifting from text generation to agent execution, and MCP standardizes client interaction with tools across multiple servers. The authors argue that tool resolution thus becomes a security boundary: metadata is not just documentation but part of the agent control plane.
- Existing gaps: The authors state that MCP requires tool names to be unique within a single server, but there is no mandatory host-wide provider-bound identity mechanism across independently operated servers; a comparative threat modeling study (reference [10]) also lists MCP's lack of provenance and identity binding verification as a security problem.
- Core observation: Familiar names, plausible descriptions, and compatible schemas can make a tool "look right" without proving its provenance; the authors hypothesize that without explicit provider binding, clients may confuse tools with similar metadata but from different providers.
- Paper contributions: Formalize the problem as an attacker-controlled tool selection optimization problem, measure it in a controlled MCP environment with a 2×3×2 factorial design (12 attacker tool conditions) on 5 LLM clients, and run robustness experiments on different names, description wording, and attacker budget.
- Threat model:
- Goal: Cause requests a user sends to a legitimate provider to be resolved to an attacker-controlled tool (tool-provider misbinding violation).
- Knowledge: The paper does not require the attacker to know client internals; the attacker can only see the effect of the metadata they advertise.
- Capability: No compromise of the host, client, LLM, legitimate server, or network channel; the attacker merely operates another MCP server visible to the same client and advertises tools with self-chosen metadata.
- Victim system: A system consisting of an MCP client/orchestrator, one legitimate MCP server, and the attacker server.
What related research exists?
Related work falls into four groups: MCP and tool advertising, LLM agent tool selection security, prompt injection and retrieval poisoning, and MCP security and software supply chain name confusion.
MCP and tool advertisement
- MCP specification (references [6]–[13]) — through JSON-RPC, clients discover tools via tools/list and invoke them via tools/call; tool advertisements include a name, a natural-language description, an input schema, and optional metadata. The paper notes that the specification requires names to be unique within a server, but there is no mandatory provider-binding identity across servers.
LLM agents and tool-selection security
- ReAct, Toolformer, API-Bank, ToolLLM, Gorilla ([14]–[18]) — study how models reason about using tools, retrieve APIs, construct arguments, and generalize to large tool sets.
- Measurement of third-party LLM application ecosystems ([19]) — found misleading application descriptions.
- Execution-layer attacks such as indirect prompt injection, malicious skills, tool-selection manipulation, cross-tool contamination, and resource abuse ([3], [20]–[24]) — the paper states that these works establish that model-driven selection can be attacked; this paper focuses on one of its consequences: when multiple providers advertise similar tools, the client may select the wrong provider.
Prompt injection and retrieval poisoning
- Prompt injection and defenses StruQ, SecAlign, AttriGuard, DataSentinel ([25]–[29]); retrieval poisoning PoisonedRAG, PANDORA, GraphRAG under Fire ([30]–[32]). The paper notes that the object of this paper is not prompts or knowledge passages but executable MCP tool advertisements; the success condition is not a target answer or jailbreak but wrong-provider resolution; and the attacker does not need imperative instructions.
MCP security and provider identity
- Research on MCP security risks ([5], [33]–[36]) and Zhao et al. ([37]) — identify risks such as malicious servers, weak provenance, and credential exposure; [37] shows that malicious instructions in external data can cause an agent to abuse a legitimate MCP tool chain.
- Protocol-level and execution-layer defense proposals ([38]–[40]) — secure MCP variants, signed tool manifests, mutual authentication, etc.; [41] addresses data access permissions, an authorization problem different from provider identity.
- Software supply-chain naming confusion ([42]–[47]) — package and update systems require artifacts to be bound to an authenticated publisher, signature, or transparency log; the authors state that MCP faces a similar binding problem in agent scenarios.
Baselines and benchmarks: The paper sets no comparison baseline method and uses no existing benchmark; the evaluation is conducted in a controlled MCP environment built by the authors (one legitimate server + one attacker server, data acquisition task).
Positioning: The authors state that this paper isolates tool-provider misbinding as provider resolution failure, systematically measures how the interaction of name, description, and schema leads to wrong-provider execution or unresolved ambiguity, and argues that provider-bound tool identity is a complement to existing defenses such as authorization, sandboxing, and user consent.
How does the paper address this problem?
Misbinding is formalized as an attacker-controlled tool selection optimization problem, using a 2×3×2 factorial design to isolate the effects of three metadata factors: name, description and schema.
Overall approach: model MCP tool-provider misbinding as an attacker-controlled tool selection problem—the attacker can only shape the provider-unbound metadata that enters the selection process, to maximize the probability that the client selects the attacker's tool; then use controlled factorial experiments to isolate the contribution of each metadata field.
Problem formalization
- Denote the visible attributes of a tool as attr(t) = {n(t), d(t), σ(t), m(t)} (name, description, schema, other metadata); if mandatory provider identity exists, add p(t).
- The legitimate tool set TL = {tL} contains only one target tool; attacker variant v is defined by three factors: Name ∈ {unrelated, exact}, Description ∈ {unrelated, copied, authority-enhanced}, Schema ∈ {incompatible, compatible}, for 12 conditions in total.
- With attacker budget r, r tools satisfying variant v are advertised, and the candidate set visible to the client is the union of the legitimate tool and the attacker tools; the main experiment uses r = 1.
- Attacker objective: max Pr[Select(q, TL ∪ TA(v,r), π) ∈ TA(v,r)], i.e., maximize the probability that an attacker tool is selected given that only advertised metadata can be controlled; the client selection policy π is treated as a black box.
Outcome categories and metrics
- Each trial is classified into one outcome: A (attacker tool selected), L (legitimate tool selected), C (clarification requested or automatic resolution refused), F (not executed).
- Main metric violation rate VR = A/(A+L+C+F); CR, FR, and LSR are also reported; when FR = 0, 1 − LSR = VR + CR.
- Resolution binding stability RBS: whether consecutive repeated executions of the same prompt resolve to the same (tool, provider) pair, normalized by N − 1 consecutive pairs; clarification and failure do not count as a resolved binding. The authors emphasize that RBS must be interpreted together with the selected provider: stable binding to the attacker is the strongest failure.
Attacker tool construction
- Name: exact directly copies the legitimate tool name (e.g., data_fetch_and_store); unrelated is constructed by task-keyword exclusion—extract a task vocabulary Ktask from the legitimate tool name, description, prompt set, and schema, and candidate names must not contain any of these tokens, must not share a meaningful substring with the legitimate name, and must retain the same snake-case format; names are generated with a fixed prompt and filtered by the rules.
- Description: unrelated describes behavior unrelated to the task; copied reuses the legitimate description verbatim; authority-enhanced preserves the original task semantics and, following a predefined template process, adds one authority cue (e.g., enterprise approval, official verification, authorization, compliance orientation) plus one safety/control cue, without introducing new functionality.
- Schema: compatible accepts the task-relevant parameters of the legitimate interface (dataset identifier and destination), so that the same call can be executed against either provider; incompatible is still a legitimate MCP tool but exposes parameters unrelated to the task. The authors state that this design separates "provider selection" from "execution compatibility."
Experimental procedure and success determination
- The system under test (SUT) = MCP client/orchestrator + legitimate server + attacker server; the legitimate tool is a dataset acquisition tool, and user prompts request execution of this benign task against a local simulated dataset and a trusted local destination.
- Fixed prompt set P = 10 (the first 5 in ordinary user wording, the last 5 in security-aware wording, explicitly emphasizing the trusted destination or expected provider); each prompt is independently repeated N = 5 times, for 50 trials per condition per model.
- All calls are logged locally, and the selected server identifier is used as the authoritative provider label (display names may be identical under same-name conditions); the experiments use no real user data, leak no data, and have no external side effects.
- Success determination: selecting a tool provided by the attacker server counts as one violation (A); VR is reported with 95% Wilson confidence intervals, and the robustness experiments additionally report the mean and sample standard deviation across metadata instances.
What experiments did the paper conduct?
5 clients × 12 conditions × 50 trials: C10/C12 had the highest violation rates, C06 showed misbinding can occur without a same-name conflict, and increased budget mainly amplified boundary conditions.
Experimental Setup
- Clients tested: GPT-5.4, Claude Sonnet 5, Gemini 2.5 Pro, Qwen3.7 Max, DeepSeek V3, 5 LLM clients in total; the paper does not report client implementations or sampling parameters.
- Environment: a controlled MCP environment with one legitimate server (the data-fetching tool data_fetch_and_store) + one attacker-controlled server; no existing benchmark.
- Conditions: the main experiment is 2 (name) × 3 (description) × 2 (schema) = 12 conditions C01–C12 (Table 1), with attacker budget r = 1.
- Sample size: 10 prompts × 5 repetitions = 50 trials per condition per model; FR = 0 for all conditions.
- Metrics: VR, CR, LSR, RBS (definitions in Question 3); VR reported with 95% Wilson confidence intervals.
- Robustness experiments: 5 different-name samples (C01–C06, 250 trials per condition per model), 5 authority-enhanced description samples D1–D5 (C05/C06/C11/C12), attacker budget r ∈ {1, 3, 5, 10} (C07–C12).
Main Results
According to Table 2 (each cell is VR/CR/RBS, 50 trials per condition, FR = 0):
Swipe horizontally to read the table
Condition GPT-5.4 Claude Sonnet 5 Gemini 2.5 Pro Qwen3.7 Max DeepSeek V3 C04 different name/copied/compatible 0.00/0.00/1.00 0.00/0.00/1.00 0.00/0.00/1.00 0.00/0.00/1.00 0.22/0.00/0.65 C06 different name/authority/compatible 0.04/0.00/0.90 0.00/0.22/0.68 0.42/0.56/0.35 0.20/0.00/1.00 0.50/0.00/0.68 C08 same name/unrelated/compatible 0.12/0.00/0.85 0.00/0.22/0.75 0.60/0.22/0.42 0.40/0.00/0.65 0.28/0.00/0.65 C10 same name/copied/compatible 1.00/0.00/1.00 0.06/0.94/0.00 0.84/0.16/0.78 1.00/0.00/1.00 1.00/0.00/1.00 C11 same name/authority/incompatible 0.04/0.00/0.90 0.00/0.24/0.68 0.02/0.52/0.23 0.18/0.00/0.82 0.56/0.00/0.57 C12 same name/authority/compatible 0.96/0.00/0.90 0.26/0.74/0.08 0.70/0.30/0.65 1.00/0.00/1.00 1.00/0.00/1.00 (For C01, C02, C03, C05, C07 and C09, VR is 0.00 or near zero for all models and is omitted; see Table 2 and Figure 2 for full data.)
- The authors state that the strongest failures came from the interaction of same-name conflict, task-aligned description and compatible schema (C10, C12), not from a single factor; C07 shows that a same-name conflict alone rarely causes attacker selection.
- The authors state that C06 shows the attack surface is not limited to name duplication: without a same-name conflict, an authority-enhanced description plus compatible schema still produced non-zero VR for four models, while Claude Sonnet 5 switched to clarification.
- Model behavior diverges: GPT-5.4, Qwen3.7 Max and DeepSeek V3 almost never clarify and directly select a tool when they fail; Claude Sonnet 5 has low direct VR but high CR under ambiguous conditions; Gemini 2.5 Pro exhibits both. The authors caution that these profiles do not constitute a general model safety ranking.
Schema Amplification Effect
- Figure 4 defines Δschema = VR_compatible − VR_incompatible; the authors state that in the C10−C09 comparison, average VR rose from 0 to about 0.78, with DeepSeek V3, GPT-5.4 and Qwen3.7 Max each rising by about 1.00 and Gemini 2.5 Pro by about 0.84.
- Claude Sonnet 5's direct VR increase is small; the authors state that it turns ambiguity into clarification rather than attacker selection; schema amplification is positive for all models, but the magnitude differs.
Robustness: Different Names and Description Wording
- Different names (Table 3, Figure 5-6): for C01, C02, C03 and C05, VR and CR are all 0 across the 5 name samples; C06 remains non-zero for DeepSeek V3 (average VR 0.696), Gemini 2.5 Pro (0.460), GPT-5.4 (0.248) and Qwen3.7 Max (0.204), while Claude Sonnet 5 only clarifies (CR 0.224). The authors state that the effect is not an artifact of a single name, but that name choice can modulate magnitude (GPT-5.4 is more sensitive to name).
- Descriptions (Table 4, Figure 7-8): C05 is all 0 across D1–D5; under C12, GPT-5.4 (0.980), Qwen3.7 Max (1.000) and DeepSeek V3 (1.000) maintain high VR across the 5 descriptions, while Claude Sonnet 5 has VR 0.256 and CR 0.744. The authors state that wording only changes magnitude, not the pattern.
Attacker Budget Sensitivity
- Table 5 (C07–C12, r ∈ {1, 3, 5, 10}): C10 and C12 are already saturated at r = 1; some aligned conditions rise with budget, e.g., GPT-5.4 in C09 rises from 0.00 to 0.34, and Qwen3.7 Max in C08 rises from 0.40 to 1.00.
- Claude Sonnet 5 rarely switches to direct selection, but its CR reaches 1.00 in C09 and C11 at r = 10; for Gemini 2.5 Pro in C10, from r = 1 to r = 10, CR drops from 0.12 to 0.06 and VR rises from 0.88 to 0.92.
- Exceptions: DeepSeek V3's VR instead decreases as budget increases in C08 (0.28→0.00) and C11 (0.56→0.00); the authors state that increased budget redistributes selection behavior rather than monotonically increasing it, and give no further explanation.
Other Ablations and Analyses
- Figure 3 model-level profiles: DeepSeek V3 has the highest average VR and almost no CR; Gemini 2.5 Pro is non-zero on all three of VR, CR and instability, which the authors call the weakest overall profile in this experiment.
- Figure 9: from r = 1 to r = 10, Claude Sonnet 5's ΔVR is small but Δ(1 − LSR) is large; the authors state that increased budget increases provider resolution failures even when it does not raise direct VR.
- Stability is not safety (Figure 2c): GPT-5.4 and others have low instability in C10/C12 because they stably bind to the attacker; Claude Sonnet 5 has high instability because it clarifies.
What are some points that can be further explored?
The authors list limitations such as a single task, only five clients, and non-exhaustive variants, and plan to extend attacker strategies; the experiments cover a controlled environment and 5 clients.
Limitations and future directions noted by the authors
- Only one controlled dataset acquisition task was evaluated, not all MCP tool categories; other tool categories may produce different rates (Section 7).
- Only five LLM clients were studied; other models, orchestrators, retrieval strategies, or confirmation designs may behave differently (Section 7).
- The metadata variants were constructed in a controlled way to isolate the effects of name, description, schema, and attacker budget, and do not exhaust all attacker strategies or real MCP deployments (Section 7).
- The authors state that these limitations affect the generality of the rates but not the core finding that "provider-unbound metadata can produce measurable wrong-provider execution or unresolved ambiguity" (Section 7).
- Future direction: extend and refine the model to cover a broader range of attacker strategies, making the model more comprehensive and better reflecting the diversity of strategies actually available to attackers (Section 7).
Experimental coverage
- The clients tested are GPT-5.4, Claude Sonnet 5, Gemini 2.5 Pro, Qwen3.7 Max, and DeepSeek V3, 5 in total (Section 5.1).
- The task is a single benign dataset acquisition task; the legitimate tool is data_fetch_and_store, using a local simulated dataset and a local destination (Section 5.1).
- The main experiment uses 50 trials per condition per model (10 prompts × 5 repetitions); the different-name robustness experiment uses 250 trials per condition per model; the attacker budget covers r ∈ {1, 3, 5, 10} (Sections 5.1, 6.1, 6.3).
- Outcome determination is based on the selected server identifier in local logs; VR is reported with 95% Wilson confidence intervals; the paper does not report the clients' specific implementations, model versions, sampling parameters, or human evaluation.
- The paper discusses defenses such as authorization, sandboxing, permission checks, and user consent, stating that they answer "whether a tool is accessible" rather than "whether a tool comes from an authorized provider," and proposes possible mechanisms such as signed tool manifests, provider-bound identifiers, schema hashing, and client enforcement, but does not implement or evaluate these defenses (Section 7).
Summarize the main content of the paper
The authors formalize and empirically measure MCP tool-provider misbinding: metadata-aligned combinations can cause clients to stably resolve requests to an attacker's tool.
A measurement-based attack paper on MCP agent security, studying "tool-provider misbinding," which occurs when tool metadata is not bound to a verified provider identity.
- Problem: MCP clients combine tools advertised by multiple servers, but names, descriptions, and schemas only help the model identify capabilities; they do not prove provenance. The attacker compromises no component and merely operates another MCP server that advertises tools with self-chosen metadata.
- Method: Misbinding is formalized as an attacker-controlled tool selection optimization problem. In a controlled MCP environment (one legitimate server + one attacker server, dataset acquisition task), a 2×3×2 factorial experiment (12 conditions) is run on name, description, and schema, with 50 trials per condition on each of 5 LLM clients, and metrics VR, CR, LSR, and RBS.
- Main results: The combination of same-name conflict + task-aligned description + compatible schema is strongest—under C10, VR is 1.00 for GPT-5.4, Qwen3.7 Max, and DeepSeek V3, 0.84 for Gemini 2.5 Pro, and only 0.06 for Claude Sonnet 5, but its CR reaches 0.94 (Table 2).
- Misbinding also occurs without same-name conflict: Under C06 (different name + authority-enhanced description + compatible schema), DeepSeek V3 VR is 0.50 and Gemini 2.5 Pro 0.42; in the different-name robustness experiment, DeepSeek V3 averages VR 0.696 across 5 name samples (Table 3).
- Attacker budget: Increasing r from 1 to 10 mainly amplifies some alignment conditions (e.g., Qwen3.7 Max rises from 0.40 to 1.00 under C08), and Claude Sonnet 5 reaches CR 1.00 under C09 and C11 at r = 10; DeepSeek V3 shows non-monotonic behavior with VR decreasing as budget increases under C08 and C11 (Table 5).
- Authors' conclusion: Tool names, descriptions, and schemas are not sufficient secure identity; model-side caution (clarification) can reduce silent attacker selection but does not produce provider binding; MCP clients need provider-bound tool identity as a complement to mechanisms such as authorization, sandboxing, and user consent.