Jul 2026 AppSec
Your agent's toolbelt is a supply chain now
Tool servers ship instructions your AI reads and your team never sees. You don't need to audit code to manage that risk. You need vendor discipline.
Read →Security · Software · AI
Small tools, built well.
An independent security lab building open-source tools. Side-project energy, production quality.
01 / TOOLS
Idle
ALPHAAn end-to-end-encrypted mobile & web client for coding agents. Run it on your machine, then monitor and drive the session from your phone or the web when you step away. Web client is live; iOS app on the way.
TypeScript · React Native · E2EE
Tide & TumbleRELEASED
A whimsical, live tide chart for US beaches.SlinkBETA
Self-hosted threat-intelligence watchtower.
Source: open-sourcing soon
02 / THINKING
Jul 2026 Reliability
The dashboard is green. The answers are wrong.
More than half of production AI agents kept perfect uptime last quarter while nearly all of them failed answer-quality checks. Your monitoring can't tell the difference. Your customers can.
Read → Jul 2026 Field notes
The prototype was the easy part
Building an app in a weekend is a solved problem. Running it safely, with real secrets and a straight face, is where the year actually goes.
Read → Jul 2026 AI governance
The chatbot in a trench coat
A lot of what's being sold as an 'AI agent' this year is last year's product with a new nameplate. Three demo questions cut through it.
Read → Jun 2026 AI governance
Nobody owns the agent inventory
Every agent in your company has a builder. Almost none has an owner. The least glamorous discipline in security is now the most urgent one.
Read →
Short takes
Jun 2026 AI governance
The approval gate is the whole point
I run coding agents most of the day, and the setting I tune most is not the model or the prompt. It's the list of actions that stop and wait for a human. Reads can auto-approve all day; anything that leaves the machine (a push, an outbound request, a package install) waits for a tap. The catch: if approving takes longer than doing the thing yourself, you'll disable the gate by Friday, and a disabled gate protects nobody. Cheap approvals are a security control. Design for them first.
May 2026 AppSec
Your AI policy covers chatbots. Your engineers run agents.
Most AI policies I get asked to review are lists of approved chat tools and rules about pasting customer data into them. Reasonable, three years ago. They have nothing to say about the agent someone wired into CI last quarter: which credentials it holds, which hosts it can reach, who reviews the code it merges while everyone is asleep. Those are answerable questions, and answering them beats drafting a new framework. Start with an inventory of what your agents can already touch. It will be longer than you expect.
Apr 2026 Threat modeling
What prompt injection actually looks like in the wild
The conference-talk version of prompt injection is a jailbreak screenshot. The version that worries me is quieter: a README in a transitive dependency that asks a coding agent to run an install script, or an HTML comment on a scraped page pointing a browser agent at an exfil URL. The user never sees any of it. The model does. Filtering inputs helps at the margins, but the surface is every string the model ingests, and that list only grows. Put the hard limits on what the agent can do, and assume it will eventually read something it obeys.
03 / CONSULTING
AI Adoption
How AI tools land inside real teams: what sticks, what stalls, and what quietly adds risk.
Secure Tooling
Building with Claude and coding agents: pipelines, guardrails, and internal tools that hold up past the demo.
Agent Guardrails
What to gate, what to log, and what to let an agent run unattended.
04 / ABOUT
Northglass Labs is an independent security and AI lab. Part security practice, part software workshop: we build the tools we want to use, then publish them.
Everything here is open source by default. Small, focused tools built with care. No venture funding, no growth metrics, no enterprise sales team. Just software that works.
The name: "North" for true north, precision and clarity. "Glass" for transparency, and for what's fragile and needs protecting.
Founded 2026 · Open source
05 / CONTACT
Building with AI, or figuring out how your team should? We like talking shop.
Or email [email protected]