PaperarXiv:2611.21577 · cs.CR · Nov 16, 2026
When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents
Zeynab Anbiaee, Sajjad Dadkhah, Ali A. Ghorbani
The factorial measurement is clean, but it covers a single toy scenario. The presentation of duplicate tools to the model is undocumented, and that is decisive for the headline exact-name results. The 'optimization' framing involves no optimization, and the proposed fix is never tested. The practical lesson, bind tools to servers in the harness, is already standard advice.
Novelty
2 out of 5Rigor
2 out of 5Reproducibility
2 out of 5Relevance to harnesses
3 out of 5
TL;DR
In a two-server MCP mock with one dataset-fetching tool, an attacker tool with the legitimate tool's exact name, its copied description and a compatible schema (C10) is selected in 0.84-1.00 of trials by four of five models (Table II). Claude Sonnet 5 mostly asks for clarification instead. The paper never says how its client presents same-named tools from different servers to the model, and that choice largely determines the headline result.
Adversarial review by Harness Hub · Nov 18, 2026 · read full text (truncated) · Suggest a correction
Original abstract (authors’ text, select to expand)
Large language model (LLM) applications increasingly act as agents that select and invoke external tools. The Model Context Protocol (MCP) standardizes this interaction by allowing clients to compose tools advertised by multiple servers. However, tool metadata used for discovery and selection is not, by itself, necessarily bound to a verified provider identity. This paper studies this gap as tool-provider misbinding: a failure in which a request intended for a legitimate provider is instead resolved to an attacker-controlled tool. The attacker does not compromise the host, LLM, legitimate server, or network channel; it only operates another MCP server that advertises tools with chosen metadata. We formulate tool provider misbinding as an attacker-controlled tool-selection optimization problem and evaluate it in a controlled MCP environment across five LLM-based clients. The results show that interactions drive misbinding. The strongest failures occur when exact-name collision, task-aligned description, and compatible schema reinforce one another, but authority-enhanced descriptions with compatible schema can also cause wrong-provider execution or unresolved provider ambiguity without name collision. Different models expose the same provider-binding weakness differently, including attacker-controlled provider selection, clarification requests, and unstable or inconsistent provider resolution.
Summary
The paper defines tool-provider misbinding: an MCP client resolves a request meant for a legitimate server to a tool advertised by a second, attacker-operated server connected to the same client. Main experiment: - A 2x3x2 factorial varies the attacker tool's name (unrelated, exact), description (unrelated, copied, authority-enhanced) and schema (incompatible, compatible). - There is one legitimate tool (data_fetch_and_store) and one benign task. - 10 prompts x 5 repetitions give 50 trials per condition per model. - Five clients are tested: GPT-5.4, Claude Sonnet 5, Gemini 2.5 Pro, Qwen3.7 Max and DeepSeek V3. - Metrics are violation rate (VR), clarification rate (CR) and a binding-stability metric (RBS). Robustness runs vary: - the unrelated-name string (5 samples); - the authority-enhanced wording (5 samples); - the number of attacker tools (r in {1,3,5,10}). Violations concentrate where exact name, aligned description and compatible schema co-occur (C10, C12). An unrelated name with an authority-enhanced description and compatible schema (C06) also gives non-zero VR for four models. The problem is framed as an optimization, but no attacker search is performed: variants are fixed templates. No defense is implemented or evaluated.
Key findings
- C10 (exact name, copied description, compatible schema), Table II. VR: GPT-5.4 1.00, Qwen3.7 Max 1.00, DeepSeek V3 1.00, Gemini 2.5 Pro 0.84, Claude Sonnet 5 0.06. Claude's CR in C10 is 0.94.
- C12 (exact name, authority-enhanced description, compatible schema), Table II. VR: GPT-5.4 0.96, Claude Sonnet 5 0.26, Gemini 2.5 Pro 0.70, Qwen3.7 Max 1.00, DeepSeek V3 1.00.
- Exact name alone (C07: unrelated description, incompatible schema) rarely leads to attacker selection. VR is 0.00-0.02 across models (Table II), although Gemini 2.5 Pro's CR in C07 is 0.20.
- Without name collision, C06 gives VR 0.42 for Gemini 2.5 Pro and 0.50 for DeepSeek V3 (Table II). Across five unrelated names, DeepSeek V3 averages 0.696 with SD 0.267 (Table III).
- Schema compatibility is the largest single contrast. Average VR rises from 0 in C09 to approximately 0.78 in C10 (Section V-B).
- Attacker budget effects are mixed (Table V). Qwen3.7 Max in C08 goes from 0.40 at r=1 to 1.00 at r=10. Claude's CR in C09 and C11 reaches 1.00 at r=10. DeepSeek V3 falls in both C08 (0.28 to 0.00) and C11 (0.56 to 0.00).
The paper at a glance
- Legitimate MCP server advertises data_fetch_and_store
- Attacker MCP server advertises variant tool (name x description x schema)
- Client LLM sees merged tool list and a user prompt (10 prompts x 5 runs)
- Outcome labeled A / L / C / F from logged server identifier
- Compute VR, CR, RBS per condition and model; robustness over names, descriptions, budget
Why it matters
For harness builders who connect several MCP servers at once (gateways, proxies, coding agents with user-installed servers), the takeaway is a design rule rather than a new attack. Tool names, descriptions and schemas are not identities, so the harness, not the model, should keep server provenance attached to every tool call. The factorial data shows that when two tools look indistinguishable, several models silently pick the attacker's. Clarification-prone models such as Claude Sonnet 5 turn the same ambiguity into user prompts. Two caveats limit how far the numbers transfer. First, the paper never documents how its client renders same-named tools from different servers. A host that namespaces or labels tools by server could change the exact-name conditions that produce the largest effects. Second, provider-bound identity is proposed but never built or tested. Read the paper as a controlled confirmation of a known risk in one toy scenario.
Issues (10)
- High severityWhere in paper: Section V-A (system under test); Section V-B 'support a protocol-level conclusion'
The client/orchestrator is never specified. The paper does not say how two tools with identical names from different servers are presented to the LLM: server prefixing, ordering, deduplication, or whether the server identity appears in context at all. These are exactly the factors behind C07-C12, which produce nearly all high VRs. Without them, the claims that failures arise under 'realistic LLM-based tool selection' and reflect a 'protocol-level' weakness are not established for deployed hosts. If the model sees no server identity, choosing the attacker or asking for clarification may be the only options it has, and the experiment then measures the test client's design.
- High severityWhere in paper: Section I contributions (last bullet); Section VII
The contributions say 'We show that provider-bound tool identity is needed to address this class of MCP tool-selection failures.' No provider-bound mechanism is implemented. No condition exposes provider identity to the client or has the client enforce it. The necessity claim is argued, not measured.
- Medium severityWhere in paper: Abstract; Section III-C Eq. (6); Section VI introduction
The abstract describes the work as an 'attacker-controlled tool-selection optimization problem.' No optimization or search over attacker metadata is run. Variants come from fixed templates and filtered generations, and Section VI says it does not seek the strongest attacker string. The reported VRs therefore say little about an attacker who actually optimizes.
- Medium severityWhere in paper: Section V-A (P=10, N=5); Section VII
Generality is narrow: - one task (dataset fetch to a local destination); - one legitimate tool; - one attacker tool in the main design; - 10 prompts and 50 trials per cell. The abstract's general statements about MCP-based agents rest on this single scenario. The limitations section concedes that rates may differ for other tool classes.
- Medium severityWhere in paper: Section V-A
Several possible confounds are not reported: - tool-list ordering (legitimate vs attacker position); - sampling temperature and decoding settings; - system prompt; - model snapshot dates. Position bias alone could account for part of the exact-name results, and no code or data link is given.
- Medium severityWhere in paper: Table II vs Tables III, IV, V; Section V-A
The same condition at r=1 reports different numbers in different tables: - C10 Gemini 2.5 Pro: VR 0.84 / CR 0.16 in Table II vs VR 0.88 / CR 0.12 in Table V. - C07 GPT-5.4: VR 0.02 in Table II vs 0.04 in Table V. - C06 GPT-5.4: VR 0.04 in Table II, a five-name mean of 0.248 (SD 0.248) in Table III, and a five-description mean of 0.036 in Table IV. The paper does not say whether these are separate runs. If they are, run-to-run variance at n=50 is comparable to effects it interprets. Table II also omits the Wilson CIs that Section V-A says are reported for VR.
- Medium severityWhere in paper: Section V-A (r>1 construction); Section VI-C; Table V
In the budget experiment with r>1, additional attacker tools use 'deterministic near-collision names when duplicate names are not permitted.' C07-C12 at r>1 are therefore not purely exact-name conditions. The budget effect is confounded with introducing near-collision names. The text also says budget 'increases pressure' and that non-monotonic drops 'do not weaken the result', yet DeepSeek V3 falls to 0.00 in C08 and C11 at r=10.
- Medium severityWhere in paper: Section III; Section II MCP security related work
The threat model assumes the user has already connected an attacker-operated MCP server. From that position, tool-description poisoning and other known attacks are already possible. The paper does not separate the incremental risk of misbinding from that of having a malicious server in the tool list at all.
- Low severityWhere in paper: Section III-D; Section V-A
The labeling procedure for clarification (C) is not described: whether automated or manual, and with what criterion. CR drives the whole characterization of Claude Sonnet 5 and Gemini 2.5 Pro.
- Low severityWhere in paper: Section V-A prompt set
Half the prompts are 'security-aware' and explicitly emphasize the trusted destination or intended provider. Results are never split by prompt type, so the effect of user-side provider hints is unknown.
Questions for the authors
- How does your client render two tools with the same name from different servers: prefixed with a server identifier, deduplicated, or passed through verbatim? Does the model see any server identity?
- Was the order of the legitimate and attacker tools in the tool list fixed or randomized, and what happens to C10/C12 VR if it is swapped?
- What temperature, system prompt and model snapshot dates were used for each client?
- Are the Table II and Table V r=1 numbers from separate runs? If so, what is the run-to-run variance per cell?
- For r>1, how many attacker tools carried the exact name versus near-collision names, and how does VR split between them?
- Do the five 'security-aware' prompts reduce VR relative to the five ordinary prompts?
- Can you add a condition where the server name or a signed provider identifier is shown to the model, or enforced by the client, to test the necessity claim?
Related in the directory
Open-source projects
Open source · Anthropic (community)
Open source · modelcontextprotocol
Open source · metatool-ai
Open source · snyk
Related papers
arXiv:2611.26595 · cs.CV · Nov 19, 2026
Code2Video uses agent-authored 3D proxies to guide video generation, enabling explicit control over structure and dynamics. However, how to construct proxies and prompts for optimal downstream quality remains underexplored, as do the capabilities and limitations of this paradigm. Through controlled experiments and blinded human evaluation, we study proxy granularity, physics simulation, and prompting. The results reveal task-dependent trade-offs: coarse proxies favor natural motion, while medium-grained proxies favor consistency. Physics simulation and proxy-aligned prompting further improve average quality. Guided by these findings, we introduce Coda-DaVinci, a task-adaptive agentic framework for Code2Video generation that selects between proxy-guided and direct generation, adapts modeling granularity to each entity, and coordinates simulation with proxy-grounded prompting. Its multi-agent harness preserves scene assets and guides repairs through visual inspection and independent review. Experiments demonstrate leading instruction following on VTR and scene consistency on R2M, improved physical plausibility over direct generation, and higher visual realism than the compared Code2Video baselines. Long-video evaluations further show gains in event structure and consistency across successive interactions, while revealing limitations in transition stability and adherence to scene settings.
Yiyan Ji, Qixun Wang +11
arXiv:2611.26553 · cs.CR · Nov 19, 2026
Agents based on large language models (LLMs) can now delete files and send payments on behalf of their users. News media reports routinely describe failures of these agents (e.g., Replit's coding agent deleting a production database), but a report often only says what went wrong and rarely (in)validates the incident or finds its root cause. Incident catalogs collect these reports, and benchmarks record whether agents fail, but neither replicates an incident nor tests the cause its report might identify. In this paper, we present ROGUE-AI (Replication of Agentic Security Events: Audit and Investigation), a multi-agent framework that closes the above gap and rebuilds each reported incident offline together with a matched control that lacks only the blamed cause, and runs agents on both. Our results suggest that ROGUE-AI is able to reproduce the reported harm in 89% of its 108 attempted replications. Agents built on five current GPT, Claude, Gemini and DeepSeek models made the reported mistake in 18% to 33% of runs, against at most 1% without the blamed cause. Of the 72 incidents with real-world harm, 62 had no underlying attacker, and in these, agents did what their task asked for but missed an ordinary hazard, such as a database shared with the live site. Replications without an attacker caused more harm than replications of prompt injection for all five models. Defenses against prompt injection look for injected instructions, but these replications contain none. Further, a warning after each tool output that names the hazard and its consequence lowered harm for all five models, whereas a note of the same length that does not mention the hazard left most of the harm in place. Built from real incidents, the ROGUE-AI replications form a reusable audit framework that vendors can rerun before each new release.
Mengxiao Wang, Mahbub Alam +2
arXiv:2611.26450 · cs.AI · Nov 19, 2026
Pretrained language models (LMs) remain unreliable on exact multi-digit arithmetic without gradient fine-tuning or external tool interpreters. We introduce closed-form, zero-gradient circuit-grafting primitives that compile deterministic computation directly into pretrained LM weights using zero training examples and zero runtime hooks, while bounding per-layer normalization drift and preserving natural-language (NL) competence. Our construction combines three analytical mechanisms: (1) two-layer Mellin-GLU quadrature over bias-free SwiGLU/GeGLU layers, turning unknown per-token RMSNorm scaling into translation along a log-spaced gate grid ($>250\times$ dynamic range) with IEEE 754 error bounds; (2) cross-token scale cancellation via signed Chebyshev Indefinite-RoPE self-sink attention and memoryless SiLU-antisymmetric GatedDeltaNet shifts; and (3) axis-aligned width extension ($d \to d+K$) with $\sqrt{d/(d+K)}$ norm compensation, Base-1000 canonical BPE emission, and differential $(+254,-256)$ tied unembeddings. An automated 5-stage compiler synthesizes 4-8 digit signed $\{+,-,\times\}$ circuits in constant $\mathsf{TC}^0$ depth alongside 3-literal Boolean (100%, $N=1{,}728$; plus 42-layer Gemma 4 E4B audits) and 5-bit Majority/Parity (100%) circuits. On 6-digit arithmetic ($N=1{,}200$, BF16), IBM Granite 4.2 3B (1.07M-1.33M active non-zeros, $\le 0.036\%$) and Alibaba Qwen 3.5 4B (2.25M-2.39M non-zeros, $\le 0.057\%$) reach 99.17% and 100.00% ALU-prefix accuracy (100% with an FP32 sidecar) and 98%-100% mid-sentence resumption, outperforming 12.8K-sample LoRA (42.2% / 61.7%, 0% inline continuation). A factorized block-sparse sidecar holds A100 memory at 7.89 / 10.65 GiB (4.60 / 8.14 GiB under INT8/INT4 + FP32 sidecar) while passing all 14/14 pre-registered $>2$K-token Tango splits ($N=4{,}392$; 0/11,134 OOD triggers).
Tanmay Sah