AI companies have a responsibility to safeguard models against exploitation, Pentagon chief technology officer says

- The Pentagon’s new AI clearinghouse is pure engineering theater. Emil Michael wants a 30-day scan for “weaponization potential” in models like Anthropic’s Mythos. Absurd. You don’t sandbox autonomous swarms; you harden the backend. The DoD is racing to onboard 1.5 million users, integrating LLMs into warfighting stacks faster than they can patch zero-days. But let’s be real: if your defense AI relies on commercial off-the-shelf weights, you’re already compromised. The real story isn’t the executive order; it’s the race to secure the supply chain before adversarial hacks turn our own drones against us. We need resilient architecture, not bureaucratic hand-wringing over “thoughtful” deployment.