Talks, slide by slide.

1. 01

   OWASP AppSec Days Portugal · Porto · 2026 · 50 slides keynote

   It's Never Been Easier to Write (In)Secure Software

   Turning AI, the fastest coder, into the safest

   Why AI made insecure code cheap, why the same shift makes secure code cheaper, and a nine-model benchmark in which the environment, not the model, set the floor.

   • 8Syntax pass rates climb to 95%; security stays in a 45 to 55% band
   • 21Our attack agent found hundreds of XSS, none in production in the safe-by-design stack
   • 36The results: 9 of 53 vibe builds executed a payload, 0 of 54 harnessed

   view → · pdf ↓ 1.4 MB

2. 02

   LocoMocoSec · Kaua'i · 2024 · 78 slides keynote

   A Recipe for Scaling (Web) Security

   Lessons from Google's Frontlines

   How Google cut XSS across hundreds of apps: fix root causes in the browser, ship frameworks that are safe by default, and patch legacy code with data and tooling.

   • 6Google's web footprint: 600+ subdomains, 900+ sensitive apps, a trillion requests
   • 36Patching in app code is easy but does not scale; patching the platform scales most
   • 48Twelve security features, each moved from opt-in to default to enforced

   view → · pdf ↓ 7.3 MB

3. 03

   SIGS — Security Interest Group Switzerland · Zürich · 2026 · 76 slides

   From Prompt Injections to Rogue Actions

   Securing AI Agents in Practice

   How prompt injection makes an agent leak data or go rogue, and how Google layers defences: a policy engine, hardened models, and sanitised output.

   • 25The lethal trifecta: sensitive data, untrusted context, the power to act
   • 34Demo: an injected review makes the travel agent exfiltrate a door code
   • 50Google's defence in depth: policy enforcement, reasoning defences, testing

   view → · pdf ↓ 5.0 MB

4. 04

   NDSS MADWeb · San Diego · 2025 · 41 slides

   Security Signals

   Making Web Security Posture Measurable at Scale

   How Google measures the security posture of 8,000+ web services from sampled HTTP traffic, then uses it to drive and track rollouts.

   • 3The scale: 8,000+ services, ~1,000 domains, trillions of requests
   • 7Architecture: the reverse proxy feeds the signal store
   • 29Four years of Trusted Types coverage, past 600 services

   view → · pdf ↓ 2.0 MB

5. 05

   SecAppDev · Leuven · 2023 · 75 slides

   Modern Security Features

   for web applications

   Four browser features that stop XSS and cross-site leaks: nonce CSP, Trusted Types, Fetch Metadata and COOP, with Google's adoption numbers from 2019 to 2023.

   • 49More than 95% of allowlist CSPs on the web are bypassable
   • 53Google 2019: CSP blocked 78% of XSS on the most sensitive domains
   • 64CSP at Google in 2023: 85% of traffic, 300+ domains, 700+ services

   view → · pdf ↓ 3.9 MB

6. 06

   Google I/O · Mountain View · 2019 · 60 slides

   Securing web apps with modern platform features

   Four browser features that stop XSS and cross-site attacks: nonce-based CSP, Trusted Types, Fetch Metadata and Cross-Origin Opener Policy.

   • 62018 bug bounty payouts: XSS alone is 35.6%
   • 19Gmail's real allowlist CSP, a full page of hosts, struck through
   • 51A complete Fetch Metadata isolation check in ten lines

   view → · pdf ↓ 3.1 MB · video ↗

7. 07

   LocoMocoSec · Kaua'i · 2019 · 56 slides

   Content Security Policy

   A successful mess between hardening and mitigation

   Google's 2018 data: a nonce-based CSP blocked 60 to 80% of reported XSS. Five policy levels, ending at nonce-only.

   • 13The case study: 7 of 9 XSS blocked on the most sensitive domains
   • 16Which sink each reported XSS used
   • 22CSP levels L1 to L4: adoption effort against covered sinks

   view → · pdf ↓ 2.7 MB

The full list, 37 talks and lectures since 2009, lives at ~/talks. Want one of these on your stage? Invite me.