It’s been a wild six weeks for frontier AI models — Mythos launched, got yanked offline by the Department of Commerce over export controls, and came back online with new guardrails, all while Five Eyes agencies warned that AI is months away from reshaping the threat landscape. This week on Safe Mode, Greg Otto talks with Armadin’s David Slater, who is building directly on top of these frontier models, creating a platform integrating AI into cybersecurity offense and defense. The two dig into what it was actually like watching America’s leading models go dark just as Chinese models — namely GLM 5.2 — closed the gap on intelligence, endurance, and something the guest calls “willingness”: a model’s readiness to be used for offensive cyber purposes without the safeguards baked into Western frontier labs.

The conversation covers why export controls and “kill switches” on U.S. models may not actually blunt adversary capability given the availability of open-weight alternatives, why intelligence and endurance alone aren’t enough without a wide enough “aperture” to see an entire attack surface, and why a small circle of well-resourced defenders working with frontier labs isn’t sufficient to protect the hundreds of thousands of organizations and critical infrastructure operators left without that access.

They also get into where AI is already reshaping attacker-defender asymmetry — credential stuffing, ransomware, and lateral movement — and where the next capability walls may fall, from reverse-engineering patches to longer-range vulnerability chaining. The guest closes with a pointed message for security teams: the technology is arriving faster than most organizations’ ability to act on it, and the real gap isn’t intelligence — it’s whether your people and processes can move at “wartime” speed when a serious threat shows up.