Staff Reports
The Pentagon’s rush to wire combat decision-making into internet‑connected chatbots just got a blunt wake‑up call. In an exclusive interview, Tyler Saltsman, CEO of EdgeRunner AI, warned that using general‑purpose, cloud‑connected models in military work could lead to catastrophic mistakes — and he pointed to a real near‑miss this spring when an AI‑assisted intelligence product almost set a U.S. operation in motion. If you care about keeping American troops and the nation safe, this should make your hair stand on end.
Pentagon AI and the “near‑miss” that should have been a red flag
The concrete example driving the alarm bells was a reported incident in which an analyst used a chatbot to synthesize open and classified signals. That AI output was folded into a routine intelligence product that falsely flagged a Chinese vessel’s cargo, and U.S. forces were readied before the error was caught. This wasn’t a spreadsheet glitch — it was an AI hallucination walking up the chain of command, looking like verified intelligence. The stakes here are obvious: a wrong turn in the kill chain can escalate into a diplomatic crisis, friendly fire, or worse.
Why on‑device, air‑gapped models deserve a real look
Saltsman argues for smaller, mission‑specific models that run on devices without internet connections. Those systems limit the attack surface, keep provenance traceable, and make it easier to enforce doctrine and human review. Full disclosure: Saltsman runs EdgeRunner, which markets on‑device military AI and has grabbed Defense contracts and private funding — so he’s selling an alternative. That doesn’t mean he’s wrong. Smart procurement should weigh both safety and vendor interest, not pretend the cloud is neutral when it’s basically a loudspeaker for risk.
Policy vacuum and the need for hard rules, not wishful thinking
Washington hasn’t passed a clear, enforceable framework for advanced AI in operations. Lawmakers talk, vendors lobby, and the Pentagon quietly expands GenAI.mil to host commercial models for millions of users. Meanwhile, analysts warn that provenance labeling, mandatory human‑in‑the‑loop checks, and audit trails are sorely needed before AI outputs feed kinetic decisions. This isn’t a philosophical debate about innovation — it’s common‑sense risk management. If leaders won’t impose guardrails, the next mistake may not be so lucky.
Conclusion: slow down, demand accountability, protect the mission
We want better decision speed and smarter tools. We do not want a chatbot making the call to send jets or board ships. The fix is simple in idea if messy in execution: limit internet‑connected models in operational workflows, require clear labels and verification for any AI‑assisted analytic product, and invest in air‑gapped, mission‑tuned alternatives where the risk is real. And if you think warnings that artificial superintelligence could be “more dangerous than nuclear weapons” are just tech‑theater, ask yourself whether you’d trust a black‑box program to decide whether lives — and peace — hang in the balance. Congress and the Pentagon should stop treating AI like an irresistible gadget and start treating it like a weapon that needs rules and human hands firmly on the trigger.
