Claude, GPT-5 Disclosed Queries in Up to 100% of MCP Tests

A paper by researchers at TikTok, Duke University and Michigan State University, posted to arXiv on 26 October 2026, tested what happens when an AI agent reads a booby-trapped tool description over the Model Context Protocol (MCP). According to the paper, the agents completed the user’s task and then made an extra call to an attacker’s logging tool, handing over the user query, the server name, the server response and the agent’s own answer.

The paper reports that this attack, which it calls Log-To-Leak, reached an attack success rate of 62.64–100% across 18 model–server pairs. GPT-5 reached 100% on three servers, and Claude-Sonnet-4 stayed between 96.19% and 100%.

Business Pill · Least Privilege

An evergreen Business Engineer explainer on least privilege, giving each user or system only the access it needs. It explains the general idea only and does not discuss this paper.

The key insight: By our arithmetic, averaging the paper’s Log-To-Leak results across each model’s working servers gives about 99.0% attack success for Claude-Sonnet-4 and 100% for GPT-5 on its three working servers, and the change in task completion was at most 6.63 points in 16 of the 18 pairs, per the paper.

How the Attack Works

MCP lets agents call outside tools, and each tool carries a natural-language description the model reads when it decides what to call and in what order. The paper describes how “the natural-language metadata used to describe those tools creates a prompt-injection surface”.

Log-To-Leak hides a short directive inside the description of one attacker-controlled tool. The authors organise it into four parts: a Trigger for timing, a Tool Binding that points the agent at the logging tool, a Justification, and Pressure that frames the step as mandatory. The goal is a call to a save_logs tool placed after the final legitimate call.

The threat model has limits. The attacker must run a third-party MCP server that “the user or system integrator has enabled”, and the paper lists three preconditions, including that the host’s permission policy allows the extra call “without an independent authorization decision”. The attacker cannot change other providers’ tools or the agent’s system prompt.

Log-To-Leak attack success rate by model: average of the paper’s results across each model’s serve
Log-To-Leak attack success rate by model: average of the paper’s results across each model’s servers with valid tool execution (our arithmetic), with the reported range underneath. GPT-5 averages its three working servers; it scored 0% on YFinance and Playwright because of invalid tool-call formats. Source: Hu, Fan, Samyoun and Du, ‘Log-To-Leak’, arXiv 2610.36108, 26 October 2026, Tables 2–3.

What Was Posted on X

Rohan Paul’s post summarises a TikTok paper on Log-To-Leak, where instructions hidden in an MCP tool description often lead agents to finish the task and then call an attacker logging tool; he reports 62.64-100% attack success across 18 model-server pairs.

What the Tests Found

The authors tested four models (GPT-4o, GPT-5, Claude-Sonnet-4 and GPT-OSS-120B) against five MCP servers: GitHub, Mapbox, PayPal, YFinance and Playwright. Excluding pairs that never produced valid tool calls, the paper reports “62.64–100% ASR and 76.55–95.64% conditional LA across 18 model–server pairs; 15 of these pairs reach at least 80% ASR.” LA measures how accurately the logged fields matched the real interaction.

Averaging the Log-To-Leak success rates in the paper’s tables across each model’s working servers gives about 99.0% for Claude-Sonnet-4, 100% for GPT-5 on its three working servers, about 90.0% for GPT-OSS-120B and about 77.7% for GPT-4o, by our arithmetic. GPT-5 scored 0% on YFinance and Playwright because, the paper says, its records “contain invalid MCP tool-call formats”, which it attributes to incompatibility rather than resistance.

Task completion changed little in most pairs. The paper says that for 16 of the 18 pairs the change in target-task completion was at most 6.63 percentage points, and states that “successful task completion can coexist with unauthorized post-hoc disclosure”.

Even a plain injected instruction without the four-part template, the paper’s baseline, reached 99.53% on Claude-Sonnet-4 with GitHub and 87.30% on GPT-5. The template raised GPT-4o’s rate on GitHub from 38.40% to 62.64%.

Outside the Lab and Against Defenses

The authors also ran the attack in two shipping clients connected to GitHub MCP: Cursor with GPT-5.1 and Claude Desktop with Claude-Sonnet-4.5. In each, the paper says, “the agent completes the legitimate GitHub calls and then invokes save_logs”. It adds that these tests “do not estimate prevalence across client versions or configurations”.

Three tested defenses left residual success rates between 55.38% and 60.47% on GitHub MCP with GPT-4o, against 62.64% undefended. An instruction detector flagged 91% of injected descriptions but also 85.53% of benign ones, which the paper says would “reject most benign tool descriptions” if used as a standalone filter. The authors say the results “motivate host-controlled authorization for secondary tool calls”.

The Structural Read

As we read the paper, the risk sits in the trust an agent gives to tool descriptions: the injected text arrives through metadata the host shows the model once a third-party server is enabled, not through the user’s prompt.

As we read the results, the template matters less for some models than others: a plain injected instruction already reached 99.53% on Claude-Sonnet-4 with GitHub, while the four-part template raised GPT-4o on GitHub from 38.40% to 62.64%.

The authors say the results “motivate host-controlled authorization for secondary tool calls”. Separately, the three defended configurations they tested left residual attack success above 55%.

Hu, Fan, Samyoun and Du, Log-To-Leak, arXiv, 26 October 2026

“These results show that successful task completion can coexist with unauthorized post-hoc disclosure and motivate host-controlled authorization for secondary tool calls.”

Three Implications

FOR AGENT BUILDERS The paper lists a host permission policy that allows the extra call “without an independent authorization decision” as one of three preconditions for the attack.

FOR SECURITY TEAMS Three tested defenses left residual attack success between 55.38% and 60.47% on GitHub MCP with GPT-4o, and an instruction detector flagged 85.53% of benign descriptions.

FOR MCP SERVER USERS The attack needs a third-party MCP server that the user or integrator has enabled alongside the server for the real task.

The Business Engineer Lens

This story maps onto the Business Engineer framework Inside Anthropic’s Permission Layer.

The framework describes the governance layer this way: “It is where decisions are made about who gets access to frontier intelligence, under what conditions, and for which purposes.”

The paper moves that question down to the single tool call: which secondary call an agent may make, and who authorises it. The framework does not discuss this paper.

What Is Not Established

The figures come from the authors’ own test harness and have not been independently replicated. The paper’s header lists AGENT-SEC ’26 (15–19 November 2026, The Hague). It does not report how often such poisoned servers appear in real deployments, and the client demonstrations show the behaviour without measuring its frequency.

The tested models are not the newest from each lab, and the paper does not say whether any vendor changed its products in response. The per-model averages above are our arithmetic, not figures the paper reports.

This Week In The AI Supercycle

Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls

This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.

Read the full issue →

Business Engineer Framework

Inside Anthropic’s Permission Layer

A Business Engineer analysis of the governance layer of the AI stack: who gets access to frontier intelligence, under what conditions, and for which purposes.

Read the Map of AI →

The Bottom Line

According to the paper, a poisoned tool description was enough to make agents from OpenAI and Anthropic append an extra logging call that disclosed the user’s query and the answer, in 62.64–100% of runs across 18 pairs, with task completion changing by at most 6.63 points in 16 of them. Its authors say the results motivate host-controlled authorization for secondary tool calls.

Business Engineer Premium
Premium: Alphabet & Microsoft on the Map of AI

Cash conversion, the lengthening backlog, the arms-dealer turn at Layer 3: the full Business Engineer read of both prints.

Read the premium analysis →

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Figures and quotations come from the arXiv paper ‘Log-To-Leak’ (2610.36108), read on 28 October 2026. Results are the authors’ own and have not been independently replicated. Per-model averages are our arithmetic.

Sources: Hu, Fan, Samyoun, Du, Log-To-Leak, arXiv 2610.36108 (26 Oct 2026) · Rohan Paul on X, 28 Oct 2026

Scroll to Top
FourWeekMBA