Schuberg Philis

Hi Network! So apparently someone thought it was a good idea to invite me to speak at a cybersecurity event about a book I wrote in 2016. Yes. 2016. When Snapchat was still cool and what I say about AI did not age well. Some examples are embarrassingly dated. But the core idea — that security should enable innovation instead of killing it — is somehow still controversial. Which says more about the industry than about the book. I'll also be announcing that I'm working on a new book: Mission-Critical Security. Same anarchist energy, better examples, no outdated AI predictions. The event is Cyber Booked and I would genuinely love some familiar faces in the room. Come support me or just come to watch me explain why I was baffled by the security industry in 2016 and am somehow still baffled in 2026.

Always a pleasure speaking to a tech crowd like that!

🤖 What if the most important engineering skill of 2026 has nothing to do with writing code? Two ideas keep reshaping how I think about AI and teams. 𝗔𝗜 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻: OpenAI's Codex team shipped a million-line app with zero human-written code. Engineers built constraints and feedback loops that made agents reliable. 𝗧𝗵𝗲 𝗥𝗮𝗹𝗽𝗵 𝗟𝗼𝗼𝗽: AI automatically writes code. Tests run automatically. Failures automatically feed back. The agent fixes itself automatically. Put them together and your team's value shifts from lines of code to problems solved. User behavior patterns changed. Revenue metrics moved upward. AI orchestration gives you scaffolding. The Ralph Loop gives you feedback. Together they move you from "we shipped 50 PRs" to "we solved 3 real problems this quarter." The teams winning with AI aren't writing better prompts. They're building better systems around the prompts. What's the one piece of infrastructure your AI workflow is still missing? #AIOrchestration #DevOpsEvolution #EngineeringProductivity #AI

If we remove the hype around the latest security LLM models, one thing is certain: finding vulnerabilities is more accessible and scalable. In other words, the likelihood of exploitation for our publicly exposed applications and devices, our software dependencies and our own developed software is higher. Thus, the "assume breach" mindset is even more relevant than before. The "when" is closer than before, and it may become "how often". How prepared are we to manage more incidents? Defense in depth is the only way to go. It matters if we are quick at patching without interruptions. It matters if we are monitoring all relevant aspects of our environment, and how long we retain our logs. It matters if we are properly managing our software dependencies and development environments. It matters if we are promptly adopting new policies and tools for the new AI landscape. It matters if our colleagues are aware of the security risks that exist in relation to their role. And so on. Most of the solutions to these problems already exist, but they must be implemented and executed well to be effective. There is also a need for speed: organizational speed. Speed to identify the gaps and address them. It is going to be intense.

When cybersecurity (gallows) humor, meets a laser cutter and engraver. Yesterday’s number was 3, but we guess we are likely going to to find out in the weekend that it was supposed to be 2 or less. #geeks #humor #security

Just held a talk with Tom Snuverink at #OpenSearchCon EU about our OpenSearch journey of the last two years of building and running clusters.

🛠️ Yesterday I said teams need to redesign the loop for agents. Here's what that actually looks like. Before any agent writes a line of code, a human defines the box it works in. One module, one concern, explicit boundaries in the prompt. The agent generates inside that box — fast, prolific, uncritical. Exactly what you'd expect. Then the human reviews. Not every line. Just three questions: → Does this introduce a new abstraction we didn't ask for? → Does it touch shared boundaries outside its scope? → Would I put this pattern in a style guide for the team? If any answer is wrong, it doesn't get reworked. It gets thrown away. Regeneration is cheap. Bad abstractions in your codebase are not. The biggest shift isn't technical. It's accepting that most agent output is disposable. Teams that treat every generation as precious end up maintaining code nobody chose. Teams that treat it as disposable keep their codebase intentional. The loop isn't write, review, ship anymore. It's scope, generate, judge, discard or ship. The discard step is the one most teams skip. It's also the one that matters most. What's your ratio of agent output you ship vs. throw away? #AgenticAI #DataEngineering #AIInProduction #MLOps #EnterpriseAI

We will always have our customers back, also when the world changes rapidly. How to handle the new reality of AI in the hands of hackers, as Claude Mythos now demonstrates

Day 2 at #AIEngineer Europe 🇬🇧 — Take aways from 2 talks. 🔥 “How Lovable Self-Improves Every Hour” — Lovable → Detect when users get stuck → Cluster failures → auto-fix → ship → Agent literally “vents to Slack” — reported bugs no one knew existed That’s real AI observability. 🧠 “Why Senior Engineers Struggle to Build AI Agents” — Google DeepMind → Text is the new state — ditch rigid schemas → Hand over control — set the goal, trust the agent → Errors are inputs — don’t restart, feed them back → Unit tests → Evals → Build agent-ready APIs — semantic, self-documenting Traditional software = aim and fire. Agents = GPS. #AIEngineering #AIEngineerEurope #Agents #DataEngineering #GoogleDeepMind #SchubergPhilis #London