Case studies
Security case studies and published tests
Five tests and one defense design, each with its method, its figures and its sources. The engineer who runs every test on this site wrote them.
The case studies
AI-built app security audit of an AI-generated codebase
Read-only check before real users arrived: 49 findings, nine of them high severity.
Prompt injection: 50 attack emails against 5 AI models
With no defense, one model called its mail tool 16 times. With a defense, 1 try in 750 succeeded.
Claude Code security review, tested on a vulnerable app
Three runs on an app with 19 known flaws: 10 found in every run, 12 in the best one.
Which AI model finds the most vulnerabilities? 5 tested
30 real vulnerabilities, each shown before and after its fix: the best model told 15 of the 30 pairs apart.
Defending AI agents against prompt injection at scale
How to stop an agent from acting on instructions hidden in an email or a web page.
A backdoor planted in a small model, and why testing can't catch it
A clean retrain cut the planted backdoor from 100% to 37%.
Have your own system tested
The same methods are behind the AI agent and chatbot security audit, the AI-built app security audit and the Web app and API security test. See all prices.