Approving an AI Dev Tool: A Security Review That Checks Instead of Trusting
Network flows, install scripts, the LLM question and the answers a vendor can't give yet. How our security pack is built to be verified on your own pilot.
November 4, 2026 · 5 min read
A development team wants a tool that reads the codebase, links it to tickets and documentation, and checks AI-written code against the rules the business validated. Before anyone installs it, someone has to approve it. Usually that’s the CISO or the infrastructure team, with a long questionnaire and a vendor reply that says “yes” to most of it.
The problem with that exchange is that a confident yes is easy to write and hard to check. We wrote our security pack the other way round. It sticks to what Lyriks states publicly about its product and installation. Where a question isn’t covered, it says so and tells you whom to ask. And it ends with checks your team runs on its own pilot, which don’t depend on anything we tell you.
Here are four ideas from it that apply to any tool you evaluate, ours included.
Ask for every network flow, by phase
“Does it phone home?” is too vague to answer well. A better request is a table: every flow, by phase, with its direction and whether it is required.
For Lyriks, installation has at most three outbound flows. HTTPS to get.lyriks.io, only if you use the one-line command. HTTPS to registry.lyriks.io to pull container images, unless images are delivered another way. Package sources for Docker, Compose or WSL2, only if they are missing on the host. In operation, there is no flow to the vendor. Users reach the web application, served at http://localhost:3000 by default, and exposing it beyond the host is your decision. Product telemetry is off by default.
The rule the pack applies to itself is the useful part: anything not listed should be treated as unknown until confirmed, and checked during your pilot with your own network monitoring.
Don’t pipe a script you haven’t read
Many security teams don’t allow curl | sh on production hosts, and they don’t have to. The scripts behind Lyriks’ one-line commands are published at get.lyriks.io/install.sh and get.lyriks.io/windows. Download one, read it, store a reviewed copy in your own repository, and run that copy. The script you read is the script the command runs.
The installer’s behavior is also documented step by step. It checks the machine, proposes what is missing, installs everything in containers including the database, runs a self-test and reports success only if the installation works, then prints the application’s address.
Pin down the LLM question in three statements
This is usually the first question about any tool that touches code and AI, and it deserves a precise answer rather than a reassuring one.
First, whose model is it? With Lyriks, it is yours: hosted, local or sovereign, with your own tokens. Lyriks doesn’t supply a model or route traffic through one of its own. Second, what does the model see? The validated spec, its rules and their sources, sent through MCP to the coding assistant you already use. That is the same material your teams already give that assistant when they work on the same codebase. Third, does the model decide anything? No. A model can write code. A deterministic engine, built in-house on graph rewriting, decides whether that code complies.
The practical consequence is that your existing LLM policy applies unchanged. If a model is already approved for your code, using it here doesn’t widen what it sees. If none is approved yet, a local model keeps the whole chain inside your perimeter.
Treat “to confirm” as information
The pack’s pre-filled questionnaire marks several answers “To confirm”: certifications, penetration test reports, encryption at rest, SSO, role-based access, backup procedure, signed images and SBOM, among others. Those questions go to contact@lyriks.io.
A “to confirm” line isn’t a yes or a no, and the pack asks you not to read it as either. We’d argue that’s what you want from every vendor. An honest gap you can chase is worth more than a confident answer you can’t verify.
Then check the rest yourself. Monitor outbound traffic from the containers during normal use and compare it with the documented flows. Connect each source with a dedicated, least-privilege account. Run a full export of the project memory and check you can read it without Lyriks. Include the container volumes in your backup scope.
What the pack contains
The full security pack runs 14 pages. Beyond what’s above, it covers the deployment model and supported platforms, a data map showing what Lyriks reads and produces, where each piece lives and who controls it, open source licensing under the AGPL and what it implies, the complete pre-filled questionnaire, and a thirteen-point checklist for your own deployment review. The Security page summarizes these commitments. To go through your own questions on a codebase, book a demo.
Go further
Security pack · free · 14 pages
Lyriks security pack, for CISOs
Deployment model, network flows, data and the LLM question, plus pre-filled answers to the usual security questionnaire.
Thank you. Download the PDF