Blog
How bomwerk builds software inventories you can hand to an auditor, and the trade-offs behind them. No hype, concrete numbers, limitations stated.
-
AI Coding Tools Are Now a Primary Attack Target. bomwerk Never Calls One.
Google's Threat Intelligence Group warned this month that AI-assisted coding tools have become a primary target for threat actors, with malware specifically designed to hide in the project directories those tools use. What the report actually found, and why a scanner that never calls a cloud AI service isn't part of that attack surface.
-
The Reporting Platform Is Live. Do You Know What You're Reporting On?
ENISA turned on the CRA Single Reporting Platform today, 11 September 2026, the same day the reporting obligation became active. Filing is genuinely simpler now: one submission, not twenty-seven. But the platform doesn't know what's inside your product, and that question was never its job to answer.
-
11 September: The 24-Hour Clock Starts. Here's What Actually Changes.
From 11 September 2026, EU manufacturers must report actively exploited vulnerabilities within 24 hours of becoming aware, and that applies to products already on the market, not just new ones. What 'aware' actually means, a worked timeline of the reporting cascade, and the three things that need to already be in place before the clock starts, because you can't build them during an incident.
-
There's No Database Server Behind bomwerk. On Purpose.
A lot of on-prem security tools quietly ask you to stand up a database before you can even run them. That's a Postgres instance to patch, a port to open, a service account to manage, on top of the tool itself. Why we store local state in SQLite instead, what that decision buys you, and where it stops being the right call.
-
Never Crash Is a Promise. Here's How We Test It.
Every parser in bomwerk returns a result instead of throwing. That's the claim from last week's post. A claim like that only means something if you can catch the moment it breaks. What a golden corpus of real repos checks that unit tests miss, one fuzz target per parser, and why undefined behavior in an archive reader is a security question, not just a correctness one.
-
A Scanner Has Three Ways to End, Not Two
Most CLI tools exit 0 or 1: fine, or not fine. That's not enough information for a CI pipeline deciding whether to fail a build. Why bomwerk uses three exit codes, what 'ran but found something wrong' actually means, and how a scanner keeps going after it hits a file it can't read.
-
Run Your Scanner Twice. Diff the Output.
Most SBOM generators produce a different file every run: fresh timestamps, fresh UUIDs, components in whatever order the filesystem handed them over. That's fine until someone asks what changed between two releases. What breaks determinism, how we pin it down, and why a scanner should never crash.
-
Your Scanner Reads Manifests. Your Firmware Doesn't Have One.
Syft and Trivy report what they can prove from a package manifest, and a typical C/C++ embedded repo doesn't have one. The three places components go missing — vendored source, static archives, forked SDKs — three commands to measure the gap on your own repo, and what heuristic identification can't tell you.