Is Your AI Better Than Their AI? Offense vs. Defense in the Age of AI
“Is your AI better than their AI?”
That was how Markus Vervier, co-founder of Persistent Security (Nemesis), opened a recent panel at Teleport, and it sums up where cybersecurity stands today. It has always been a high stakes chess match, but lately it feels more like a high frequency trading floor, with attackers and defenders both plugging frontier AI models straight into the network. When both sides have access to the same caliber of models, where does the edge come from?
Ben Arent (Director of Product at Teleport) sat down with Markus and Rob Picard (Director of Security at Teleport) to get an overview of what happens when offense and defense collide in the age of AI.
The Scaling Shift: From Human to High Frequency Hacking
Back in the day, finding deep infrastructure flaws or pulling off complex code execution required specialized human knowledge. Today? Large Language Models (LLMs) and AI agents have read every spec, every writeup and every CTF (Capture the Flag) challenge ever published.
As Markus pointed out, AI allows attackers to scale their capabilities instantly. But here is the catch: infrastructure security matters more than ever. When automated swarms can chain together five low severity bugs into a critical breach at lightning speed, your containment strategy is your last line of defense.
The Hugging Face & OpenAI Incident
If you need proof that this future is already here, look no further than the bizarre incident where Hugging Face suspected a nation state attack, only for OpenAI to step up a few days later and say, “Oops, sorry, that was actually our autonomous model testing things out.”
It highlights a harsh reality: AI agents are getting exceptionally good at chaining vulnerabilities autonomously. When the boundaries between human intent, automated models, and infrastructure blur, border defense security alone won’t save you.
Frontier Models: Reinventing the Wheel vs. True Innovation
Everyone is talking about mythic frontier models, but are they actually inventing new attack vectors? Not quite yet.
As Rob noted, someone had to invent SQL injection or return oriented programming (ROP) from scratch. Current frontier models aren’t quite inventing completely novel classes of math or computer science breakthroughs out of thin air. Instead, they are acting as hyper efficient vulnerability scanners and master orchestrators, chaining known concepts together faster than any human team could map.
Markus adds that general purpose frontier models hit a wall because they try to be good at everything. The real evolution will live in specialized multi agent swarms, models tailored explicitly to architecting and defending specific environments.
The Nemesis Perspective: Moving Past Traditional Pentesting
If we use AI simply to scale up traditional penetration testing, we are just drowning ourselves in a sea of false positives. Instead, the real win is shifting toward property based testing and automated threat modeling.
Imagine defining strict infrastructure invariants like “Database X is mathematically and structurally incapable of communicating with the outside world” and letting AI verify those properties continuously. Rather than just throwing random payloads, Nemesis uses AI to prove that the security posture actually holds up under pressure.
So, Is Your AI Better Than Their AI?
When attackers and defenders have access to the exact same caliber of AI models, the answer is less about the model and more about how you use it. It ultimately boils down to three things:
Speed: Who can detect and patch faster than the automated script running on the other side.
Operational Efficiency: Having a smarter, leaner model architecture.
The Human Filter: You still need human experts to direct the AI, manage the guardrails, and focus on high impact risk reduction.
Final Thoughts
Cyber Security is facing its most exciting era yet. As Rob put it, security professionals lean heavily toward cynicism, but treating AI as a collaborator rather than a magic wand changes everything.
Don’t use AI to scale up broken old workflows. Use it to build smarter systems, tighten your infrastructure properties, and stay one step ahead.
