You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$9,000 left to reach the funding goal.
This project extends an evidence-driven browser security methodology — already applied in two published Zenodo reports on Chrome 152 and Edge 152 (~1.94 million fuzzing executions, no crashes, hash-verified evidence manifests) — to the AI-agent features now shipping in Chromium-based browsers.
Working independently, I will test how prompt injection and context boundaries behave when AI agents interact with web content, building directly on an open thread from the Edge report: a data-minimization observation in Copilot. The project does not assume a vulnerability exists; it is designed to determine what can be demonstrated under controlled conditions and state clearly what cannot.
The deliverable is a reproducible technical report, published and archived within 8–10 weeks. Success is measured by the quality and verifiability of the evidence — a well-documented negative result counts as much as a confirmed finding.
This project builds on two research reports published on Zenodo, covering Chrome 152 and Edge 152. Across the work, I conducted approximately 1.94 million fuzzing executions without crashes. The methodology emphasizes controlled and reproducible experiments, including frozen browser builds, disposable profiles, and hash-verified manifests to preserve experimental integrity and make results independently verifiable.
One open research thread concerns a data-minimization observation involving Copilot. I want to examine this question further while extending the same disciplined methodology to browser surfaces involving agentic AI, particularly prompt-injection risks and context boundaries. The objective is not to assume that a vulnerability exists, but to determine what can be demonstrated under controlled conditions, distinguish observed behavior from interpretation, and identify where the available evidence leaves questions unresolved.
I will approach the work incrementally: establish the tested builds and experimental conditions, define repeatable test cases, preserve the relevant artifacts and logs, and document both positive and negative results. Any claims will be tied to the evidence obtained, with limitations stated explicitly.
The primary deliverable will be a reproducible technical report, published and archived within 8–10 weeks. It will document the research questions, test environment, methodology, results, supporting evidence, and limitations in sufficient detail for another researcher to understand and, where feasible, repeat the experiments.
Success will be measured by the quality and verifiability of the evidence, not simply by whether a vulnerability is found. My previous browser research has shown the value of documenting what extensive testing did and did not establish. A well-supported negative result—one that clearly defines what was tested, under which conditions, and what conclusions the evidence permits—will count as a successful outcome. If a vulnerability or security weakness is identified, the report will distinguish the reproduced behavior from any broader claims about impact or exploitability.
The goal is to produce a technically defensible, independently reviewable contribution, whether the findings confirm a security weakness, narrow the conditions under which a suspected risk could occur, or establish that the tested scenarios did not demonstrate the hypothesized issue.
At the minimum funding level ($4,500), the project covers core testing costs (API/model access, compute, and tooling) with a reduced time allocation, narrowing the scope to the most tractable test cases. The full $9,000 goal funds the complete 8–10 week sprint as scoped below
$4,500 time, $1,200 API/models, $1,000 hardware/compute, $800 software/tooling, $1,500 contingency.
I am working independently on this project, without a formal team or institutional backing. My research record includes two Zenodo deposits (https://doi.org/10.5281/zenodo.22877622 (Chrome 152)
https://doi.org/10.5281/zenodo.22681660 (Edge 152)), my ORCID profile (https://orcid.org/0009-0008-4388-1836), work available through Research Square, and a manuscript currently in its third round of review at AI and Ethics. Together, these provide evidence of my research practice and methodological rigor — including ~1.94 million fuzzing executions across the Chrome and Edge investigations with no crashes or sanitizer artifacts, and a specific open finding (a data-minimization observation in Edge’s Copilot assistant) that this project directly extends. I believe this track record transfers directly to the proposed work. I am building the work step by step, with an emphasis on reproducibility, transparent documentation, and clearly stated limitations.
The main risk is that agentic browsing capabilities may not yet be mature enough, or sufficiently available, to support systematic testing within the project timeline. If that happens, I will adjust the scope to focus on the capabilities that are actually available and can be tested reliably. I will document the limitations, explain what could and could not be established, and report the results honestly, including negative or inconclusive findings. I would rather deliver a narrower, well-supported result than force a finding that the evidence does not justify.
I have raised $0 in the past 12 months. I currently have an application under evaluation at EA Funds, submitted for a broader research program that does not depend on this project. It has been under evaluation since October 2. On October 25, BlueDot declined a smaller funding application for this same project, without providing specific feedback.
I am being transparent about my funding situation because I want the project to be assessed on its actual merits and current resources, rather than on funding that has not yet materialized.