Security standards for teams building with AI coding assistants

TLDR/BLUF

What this is. Standards you can adopt, fork, or hand to an auditor. Each ships as markdown and as a Word file, written to be copied whole into your own repository and edited.

Why it matters. A control you claim has to be one you can show firing. Several of these documents exist to make that difference visible.

Benefits. Adoption hands you a drafted standard rather than a blank page, MIT licensed, yours to rename and edit. The set also publishes the figures that did not survive checking, named as claims to stop repeating next to the ones that held.

Costs. Adoption is per document and each says its own price; nothing here is a product, a scanner, or a service, and nothing is sold. You end up maintaining the text yourself, and citing the model document’s permanent rule identifiers in your own deviations register.

Not for you if you want a certification, an attestation, a maturity score or a tool recommendation. What these are not says why none is offered.

Where to start. Which security standards actually apply to you routes on your situation rather than on the documents. The standards index has one row each. The CISO summary is the argument in two pages.


Where each reader goes

You are Start at
Deciding whether any of this reaches you Which security standards actually apply to you
Briefing an executive, or asking for funding The CISO summary
Writing or forking your own standard Secure development, the model, with stable rule identifiers
Deciding how much AI-written code a human reads Human review of code
Setting expectations for AI-assisted work AI-assisted development
Judging whether a body of code is any good Code quality
Managing code you did not write Dependency and artifact integrity
Working out which rules a pipeline can enforce Automated compliance in CI
Selling software, and being asked for evidence What to have ready when a buyer asks
Adopting any of it, in an order that works How to adopt these

What these are not

No certification, and no attestation. None is offered and none is implied. These set a bar to hold, not a badge to present, and no maturity score comes with them.

No tool recommendation. No scanner, format or assessor is recommended anywhere in the set.

No legal force. Whether any of this binds your organization is decided by a clause number and by counsel, never by a web page. What these documents do is tell you which questions are real and what a good answer looks like.


What makes these different from a checklist

One claim runs through every document: a control that reports success is indistinguishable from a control that never ran, so a claim is worth what its receipt is worth and nothing more. That is why:

  • secure development carries permanent rule identifiers you can cite in your own deviations register
  • every rule has an evidence column
  • automated compliance in CI turns on one question – does failing it actually stop the change

The set also publishes what did not survive checking. Widely repeated figures about machine-written code are named in code quality as claims to stop repeating, next to the ones that held. A document that only published its wins would be asking for trust it had not earned.


Running an assessment, and the pipeline it runs in

Two method documents sit beside the standards rather than inside them. Running a large security-standard assessment with AI agents is what was learned scoring several hundred requirements with agent sessions, and it is candid that the obvious way to split that work is the wrong one. CI and standards for agent-written code covers getting parallel work through a pipeline without believing things that are not true.


Where these came from

They were written alongside claude-multisession, a toolkit for running several agent sessions against one repository, and that toolkit remains their worked demonstration: it applies the same claim to concurrency, where every control fires on purpose and prints what it examined. The documents outgrew it and now live here.

Everything is MIT licensed. Copy it, fork it, put your own name on it.