Writing on the agentic web Roughly once a month
Cryptography, local inference, and what it actually takes to build a browser from first principles. Written by the people doing the work.
Photo: Hassan Pasha / UnsplashA browser that can navigate, fill forms, and submit them is a qualitatively different security problem from one that only renders.
Here is how we think about prompt injection, data exfiltration, and the permission model that has to address both at once.
Read the post →Nation-state adversaries are collecting encrypted traffic now, planning to decrypt it once quantum computers arrive. For anything sensitive, the window to act is shorter than most teams think.
We ran a 30-day network trace on the three major browsers at default settings. The results were worse than we expected, and clarifying about what “opt-out” means in practice.
We benchmarked local inference for the Dart Agent on M3, M4, and AMD AI-MAX hardware. Here is what is actually usable for agentic workloads today.
Every browser that proxies AI inference has a structural conflict between its margin and your wallet. Here is why BYOM is the only honest pricing model for a browser with an AI layer.
How we use TPM 2.0 and the Apple Secure Enclave to generate non-exportable identity keys — and why that changes the threat model for browser-based authentication.
A walkthrough of the patches we apply to the Chromium tree: what stays, what goes, and the engineering trade-offs in removing tightly-coupled Google services.
New writing lands roughly once a month: engineering notes, benchmarks, and the occasional argument. No product announcements dressed up as insight. If there is something you want us to dig into, write to team@dartbrowser.com — a person reads every message.
