Quantitative AI Security
Security, measured.
Every Sunday at 5pm ET, I take one AI security claim, from a paper, a vendor, or my own work, and check it with four questions. You get the code, the numbers, and a verdict.
Free when you sign up: The QAS Rubric, a 14-page guide to checking any AI security claim.
What you get every Sunday
One claim, checked.
Can the claim be validated? Where is the code? Does it work outside the lab, at scale? Each check ends with one of six verdicts: holds, doesn't hold, unvalidatable as published, no code, lab-only, or deployable.
The receipts.
Code you can run and numbers you can recheck. When my own work fails the same test, you see that too.
A few links worth your time.
Filtered hard. One line on why each one matters.
Free with signup
The QAS Rubric: four questions for any AI security claim
- Can I validate this claim?
- Can it even be validated?
- Where is the code proof?
- Is it feasible in a real security environment, at scale?
Inside: the red flags for each question, the six verdicts, four worked verdicts on real 2026 papers and vendor reports, the same test applied to my own tool, and a one-page checklist to print. Use it the next time a vendor shows you a benchmark.
Recent issues
Get the Sunday issue
Written by James Bower. 10+ years in ML-driven cyber ops and threat intel.