Mirage
Mirage shows what an AI coding agent read on a web page that a person could not see,and blocks the actions that hidden text tries to trigger
About the project
Pages can hide instructions that people never see but AI coding agents read and follow: off-screen text, zero-size fonts, HTML comments, JSON-LD, invisible Unicode, or a different page served only to AI user agents.
Mirage loads every page twice: once as the agent gets it (raw HTML with an AI user agent) and once as a person sees it (real Chrome via Bright Data). Text the agent got but a person can’t see is flagged, shown in an X-ray screenshot, and any action the agent tries because of it (installs, secret reads, payments, outbound requests) is blocked and traced back to the page it came from.
Demo: a Gemini coding agent fixes a checkout bug while reading 5 pages, 4 with hidden instructions. It tries to install an unknown package and print the environment; Mirage blocks both. Re-run without the hidden text and the same agent fixes the bug safely.
Measured: 11 of 11 planted attacks caught, 0 of 5 clean test pages flagged, 0 of 64 real developer pages flagged.
Built with
AI tools
Prize Categories Entered
- Submitted

Best Use of Gemini API
Gemini 3 Flash (preview) powers the coding agent through function calling (its tool calls are what Mirage guards) and writes the one-sentence verdict for each finding, with structured JSON output.
- Submitted

Best Use of Tiger Data
Tiger Cloud (Postgres + TimescaleDB) stores runs, steps, findings, taints and saved scans; the Mirage Index uses a hypertable with a continuous aggregate for hourly trends.
- Submitted
Best Use of Bright Data
the Scraping Browser renders each page in real Chrome to get what a person sees (visibility check plus normal and X-ray screenshots), with Web Unlocker as a fast fallback.
- Submitted
Best Demo Video Using HyperFrames
To create demo