QueryShadow

QueryShadow is a self-reported privacy tool for AI agents that visualizes and analyzes the cumulative exposure of an agent's search traces — the ordered sequence of queries it makes — before those queries are sent to external sources. The author states that it was built as a prototype for the OpenAI 2026 hackathon, with no secret data or revenue evidence.

The product is described as a browser-local, no-login tool that replays agent searches and reconstructs what an outside observer could infer from them. It offers a JSON export of mitigation receipts and aims to make privacy risks visible in the context of agent trace order, not individual prompts.

The description states that QueryShadow:

• replays an AI agent’s outbound searches;
• treats the ordered sequence of queries as the unit of privacy risk;
• shows cumulative exposure rising across the trace;
• reconstructs what a plausible outside observer could infer;
• links inferences to literal bridge evidence;
• rewrites only identities, values, dates, locations, percentages, and other revealing details;
• estimates retained research utility and residual exposure;
• exports a transparent JSON mitigation receipt.

The judge path is described as instant, browser-local, no-login, and needs no API key. It is built with React 19, TypeScript, Vite, Cloudflare Workers, OpenAI Sites, and GPT-5.6.

Inference The product appears to be a static analysis tool for reviewing agent traces in a synthetic or local context, not an active enforcement system.

The author states that QueryShadow was built to make “cumulative mosaic leakage visible before an agent sends the trail.” It is positioned as a tool that addresses privacy risks at the trace level — not individual prompts or tool calls — and aims to close the loop from reconstruction to evidence to safer query plan to residual-risk receipt.

It claims to be different from other privacy controls, which inspect one prompt, file, or tool call at a time. Instead, it treats the ordered trace as the security boundary.

Inference The positioning is that of a privacy review/debugging tool for AI agents, not an enforcement or blocking system.

The author states that QueryShadow was built to address risks in three synthetic scenarios: M&A diligence, clinical research, and unreleased product launch.

There is no evidence of actual customers or target accounts beyond these use cases. The tool is described as a prototype for a hackathon.

Inference The intended audience is likely AI agent developers or privacy-conscious organizations using agents in sensitive contexts — but no real-world customer data is provided.

There is no evidence of pricing, monetization, or business model. The product is described as a prototype for a hackathon and is not presented as a commercial offering.

Inference No business model or pricing information is evidenced.

The tool is built with:

• React 19
• TypeScript
• Vite
• Cloudflare Workers
• OpenAI Sites
• GPT-5.6 and Codex

It uses a deterministic analyzer that extracts signal classes, accumulates context in query order, identifies bridge-entity joins, and generates minimal generalizations.

The interface is described as responsive, with trace import and audit export capabilities. It is browser-local and does not transmit data.

Inference The tool is technically feasible as a prototype but lacks live agent integration or production-grade delivery mechanisms.

The project is described as a hackathon submission (OpenAI 2026). It includes:

• Three synthetic scenarios;
• A responsive UI;
• Trace import and audit export;
• Transparent limitations;
• Deterministic tests;
• A sub-three-minute narrated demo.

There is no evidence of real-world adoption, user base, or revenue. The tool is described as a “review debugger,” not a live system.

Inference No traction or maturity signals are evidenced beyond the prototype nature of the product.

The author states that most privacy controls inspect one prompt, file, or tool call at a time. QueryShadow is positioned to treat the cumulative ordered trace as the security boundary.

There is no evidence of existing competitive products or market positioning in the description.

Inference The competitive context is not evidenced, but the product appears to address a gap in privacy controls for AI agent traces.

• The tool is described as a prototype with synthetic scenarios and no live agent integration.
• It is browser-local and does not transmit data — this may limit its utility for real-world deployment.
• No evidence of revenue, customers, or commercial traction.
• The author explicitly states that it is a “review debugger,” not a system that enforces or claims actual inference by external providers.

Inference The main risk is that the product remains a prototype and has no demonstrated path to commercialization or real-world adoption.

1. What are the technical limitations of the current prototype, and how would they be addressed in a production version?
2. Are there any plans for live agent-trace adapters or integration with real AI agents?
3. How does the tool handle edge cases or ambiguous inference scenarios?
4. Is there any evidence of interest from potential customers or partners?
5. What is the roadmap for moving from prototype to commercial product?