Insights

Staff send customer records, code and contracts to AI tools every day, through more paths than the chat box. Where the data goes, what happens to it there, and the controls that work without a ban.

The 2026 OWASP list says no reliable prevention for prompt injection exists. What an AI gateway can still do about the ten risks, which part stays with your application, and what to ask a vendor.

Two critical flaws in PHP itself allow unauthenticated command execution. The fixed versions, why php -v misleads on distro builds, interim settings, signs of attack in your logs, and a manager's checklist.

A policy an engineering lead can adopt this week — what must never be pasted into an assistant, how to review the code and packages it produces, and a step-by-step rollout.

Measure a web application firewall profile by profile — detection rate per attack class, false positives on real traffic, the common bypass families, and the tuning loop that follows.

Persian text can carry the same instruction in a form your filter does not recognise and the model reads perfectly. How to test for it, and what to fix.

What changes when the application under test runs on a language model, the data-flow map that scopes the test, the eight areas to cover with examples, and how to report an attack that is not reliable.

The sections a useful penetration test report has, what each one is for, and the questions to ask before you accept a deliverable that is missing them.

The two are bought for different reasons and priced differently. How they differ in method, depth and output, and a simple rule for which one to order this quarter.