Rendering QA agent for 150M+ readers of Lao · Thai · Khmer · Burmese

Your site says it speaks ພາສາລາວ. Does it?

GlyphSweep finds clipped tone marks, missing fonts and broken syllables, writes a CSS fix, and proves it by re-rendering. Every page it touches opens in a throwaway gVisor sandbox on Vultr.

A coffee shop, in whichever script you picked, that tries to steal cloud credentials and hijack the AI. Watch it fail.
Real site: laocoffee.org/lao/ → A real Lao coffee site, swept live: bare-phone OCR error 94% → 7%, score 50 → 74.
Disposable gVisor sandbox per stageZero secrets insideVultr Serverless InferenceRed-teamed: results ↓
144

live websites measured in Lao, Thai, Khmer and Burmese: government, banks, universities, telecoms, news

51%

have a visible rendering problem: clipped marks, broken syllables or no font for the script

45%

never deliver a font for their own language, so text depends on the visitor's phone

49%

declare the wrong lang, so screen readers read it in the wrong voice

ລາວLao50%82 sites
ไทยThai38%24 sites
ខ្មែរKhmer63%19 sites
မြန်မာBurmese74%19 sites

Share of sites with a visible rendering problem, by script. Measured with GlyphSweep's own detectors at desktop + phone width from a Vultr VM, Nov 2026. Method, limits & data →

How the agent works

Nothing is trusted. Everything is measured.

The model proposes. Measurements decide. A human releases.

01 · planScope

Validates the URL and builds a per-job egress allowlist.

02 · dispatchSandbox render

Chromium in gVisor runs deterministic detectors on real glyph ink.

03 · bare deviceTofu → number

Re-renders with no script font; OCR measures what a phone shows.

04 · verifyVision check

A vision model on Vultr confirms what a reader actually sees.

05 · fixSanitized CSS

The model writes a patch and translations; a sanitizer strips anything risky.

06 · proveFresh sandbox

Patches compete in new sandboxes. Only the best measured one ships.

Live agent

Run it. Watch every sandbox live and die.

Agent timeline

Plan → sandbox → verify with vision → fix → prove in a fresh sandbox.

Run a sweep, or the hostile demo, to see results.
Red team · attack our own sandbox

We tried to break out. Run it yourself.

–
Five fresh gVisor sandboxes attack themselves: hostile pages, code execution, memory bomb, fork bomb, hang. Verdicts come from the egress proxy's log and Docker, not from the sandbox's own word.
Loading the latest run…

Badge for your README: ![red-team](https://45-77-254-54.sslip.io/api/redteam/badge.svg) · attack pages: control/demo/redteam