Skip to main content

August 20, 2026

Highlights​

This is the release where App Builder learns to try to break your app, on purpose, before anyone else does.

  • Builders: your app can hire its own red team. Pick a depth, start an assessment, and go back to building. App Builder reads the code, then (if you want) actually attacks the running app. Findings are mapped to OWASP Top 10 and OWASP LLM Top 10, the lists security teams already use for web apps and AI apps. Confirmed issues come back with proof, a suggested fix, and a button to send them straight to the builder.
  • Everyone on the team: security is no longer a vibe. Every completed run gets an A-F grade and a score out of 100, using the same kind of severity scoring (CVSS) that professional assessments use. That grade shows on the app gallery and in the deploy queue, so you can see posture at a glance instead of hoping the last publish was fine.
  • Org admins: you set the rules. Decide who can run assessments, keep Full Red Team to admins, and open results from an email link. You get the deep scan when you want it, without putting the expensive tier in everyone's menu.

We are rolling this out progressively, organization by organization. If you do not see Run security assessment yet, it has not reached you. There is a fuller walkthrough below, plus a few everyday reliability improvements.

New features​

App Builder can now try to break your app​

Shipping an app used to mean trusting that nothing obvious was wrong. That is a lot of trust.

App Builder can now run a security assessment on demand, against the app you are actually building. It is not a static checklist you ignore. Deeper tiers run the app and attack it, then only report issues they can reproduce.

This is built to read like a real assessment, not a linter with nicer colors. Findings are classified against OWASP Top 10 (2021) (the standard list of web app risks) and OWASP LLM Top 10 (the same idea for AI apps: prompt injection, tool abuse, leaking secrets). Severity uses CVSS v3.1, so a High means what a High means everywhere else. The scan aims at OWASP ASVS Level 2, the bar most apps that handle user data are held to. That is a coverage target, not a certification.

Choose how far to go:

  • Quick Scan (~1-2 min): a careful read of code and dependencies. Hardcoded secrets, known-vulnerable libraries, missing login checks. No live attacking.
  • Break-in Test (~5-10 min): everything above, plus it boots the app and tries the attacks real people use: injection, cross-site scripting, broken logins, data you should not be able to see. If it cannot reproduce it, you do not see a fake finding.
  • Full Red Team (~20-40 min): the works, including attacks unique to AI apps: prompt injection, tool and connector abuse, chained exploits, and multiple rounds until nothing new turns up. Org admins can keep this tier to themselves.

What you walk away with:

  • A grade from A (safer) to F (serious issues) and a score out of 100. A confirmed critical finding caps the grade at D, so a pretty dashboard cannot hide a real hole.
  • A report with severity, proof, and how to fix each issue, grouped the way OWASP groups them. Confirmed means we reproduced it. Unverified means it is worth a look, but it was not exploited live.
  • Fix selected: send one finding or a handful back into the builder, in the same place you already work.
  • A PDF you can download and share, with those same standards on the cover. Emails deep-link to the results, so you are not hunting through the UI for last Tuesday's run.

Assessments keep going if you close the window. Start one from an app's actions, or from the Security view for a list across apps. Gallery cards show a security-grade seal after a run, so the next person does not have to guess.

Admins stay in control. In organization policy you can allow or block members from running assessments, and separately allow or block Full Red Team. Members never change the assessment model. The default is a strong auto pick.

The product is ready. We are enabling it progressively. Look for Run security assessment. If it is not there yet, more communications will follow as it reaches your organization.

Improvements​

  • App gallery cards are easier to act on, with a menu on each card and that security-grade seal when an assessment has run.
  • Gemini 3.7 Flash is in the model catalog.
  • File uploads use private, time-limited access.
  • Knowledge links for Jira and Drive open more reliably.
  • App Builder previews keep a path bar, so you can see the route you are looking at.

Bug fixes​

General reliability improvements across App Builder security assessments, file handling, and knowledge. Together they should make scans and day-to-day use feel steadier.

⌘J