Theodore Tso, Thank you for reporting this. We will be improving the browser based on your findings. By performing this security review and publishing the findings, you are helping make this fine web browser more secure. I would welcome a followup security review in some time so that we can verify that all the issues have been resolved.
Sincerely, Andreas Røsdal On Fri, Oct 23, 2026 at 3:04 AM Theodore Tso <[email protected]> wrote: > On Tue, Oct 20, 2026 at 08:48:52AM -0500, Andreas Rųsdal wrote: > > Thank you for your feedback. I would appreciate feedback from more people > > to get a more nuanced and neutral view, > > preferably from someone who actually works on real software development > in > > 2026, someone with education in computer science, > > also considering that software development has changed with the > > introduction of AI llms. > > Normally I tend not to do the full brag sheet, since I believe that > technical arguments should not be resolved using argumentation by > credentials / authority. But since you asked.... (and for more > details feel free to look at the Wikipedia entry describing my humble > self, although it is a bit outdated these days :-). > > I graduated from MIT in 1990 with a degree in Computer Science and > Electrical Engineering, with a specialization in Computer Science and > a minor in Economics. I was the first North American Linux Kernel > developer, and am currently serving as the ext4 file system > maintainer, so I've been working on the Linux kernel for 35 years. I > have also served on the Internet Engineering Task Force, as one of the > working group chairs for the IP Security Working Group, and was one of > the members of the Security Area Directorate. > > In my day job, I am one of the senior technical leads on my employer's > data center kernel team, where we have teams working on agentic-aided > kernel review (Sashiko), agentic cherry-picking and merge conflict > resolution. agentic-assisted kernel debugging, and agentic assisted > kernel live-patching. I personally use Large Language Models in my > work. For example, just recently I created a prompt[1] to create > LLM-generated Kernel Unit tests[2]. > > [1] > https://lore.kernel.org/all/[email protected]/2-writing-kunit-tests.md > [2] https://lore.kernel.org/all/[email protected]/ > > So with my credentials established... > > 1. I agree with the concerns expressed by others, most notably by > Gard Spreemann. > > 2. It seems strange that you have recently established a "No AI" > policy starting in October 2026, given that previous to that > pretty much all of the code base was created using AI. Why is > that? (There was also a commit that was authored by Claude > *after* the no AI policy, although it appeared to be deleting > tests.) > > 3. I didn't see any tests in the repo, which was strange. Maybe they > are located elsewhere? I *hope* they were used as part of the > agentic coding loop, but why did you recently delete the > cute-tests directory? > > 4. The use of C instead of Rust for a software package where you are > starting from scratch and which was vibe coded was strange. Rust > allows the compiler to find many bugs autpomatically, which is > really useful for agentic code harnesses. It's especially > interesting when you try to claim that it is more secure than all > other browser code bases, apparently because AI(tm) and because it > was coded from scratch, since C is susceptible to use-after-free > bugs (for a not-hypothetical example). > > 5. Because I'm a paranoid SOB, and given your hyperbolic security > claims, I decided to run a little experiment. So I asked a > frontier model developed by my employer, "Look for potential > security issues in the Northstar Web Browser. Not just whether it > adheres to the security model found in SECURITY.md but also look > for the possibility of deliberate back doors that were introduced > by the authors." > > 5 subagents, 4 million uncached input tokens, 14.5 million cached > input tokens, and 217k output tokens later, I received the report > which I've attached as FINDINGS.md. It is also available here[3] > which may be easier for people to read. > > [3] > https://docs.google.com/document/d/1sNxSNDpsaLXO-WsKpeKCWajH1nfCkjLg/edit#heading=h.kkp50pksh0c7 > > The short version is that multiple severe security vulnerabilities > were found, including multiple use-after-free that could be used to > attack a user using the Northstar browser. The "Arbitrary Local > File Disclosure from JS" was also somewhat amusing given the > assertions that this was more secure than Firefox and Chrome. As a > result, I would **not** recommend that any user depend on Northstar > Browser if they cared about their security. And perhaps the Debian > FTP Masters might want to take this into consideration when > deciding whether or not to accept this package in its current form, > and whether the Debian release team would be willing to release it > as part of a future Debian Stable release. > > I'd also recommend that you consider rescinding your "No AI policy" > and get help from your AI agents to fix the 8 Critical, 7 High, 3 > Medium, and 1 Low severity vulnerabilities identified in > FINDINGS.md[3]. > > Cheers, > > - Ted >
