Hot take: DARPA's AI flew an F-16. The switch that took it back is the real story.

Hot take: DARPA's AI didn't fly that F-16. The switch that handed it back to a human did the actual work.

Here's what happened, stripped of the headline. DARPA and the US Air Force flew a modified F-16 at Eglin Air Force Base under the VENOM program — Viper Experimentation and Next-gen Operations Model. A human test pilot took off, flipped a switch, and handed flight control to an AI agent mid-air. The pilot stayed in the cockpit the entire time, ready to take control back, and did exactly that to land the jet.

That's not "no pilot required." That's a fielded fighter jet with an aftermarket autonomy kit and a kill switch. And that distinction is the whole story.

I've spent years building perception and control systems for defence deployments — real-time threat detection running at 94% precision, the kind of system where a false negative isn't a bad demo, it's a casualty. Every one of those systems shipped with the same non-negotiable: a human authority layer that can override the model in under a second, with zero ambiguity about who's in control at any given moment.

What DARPA built here is the same pattern at fighter-jet scale. The VENOM kit doesn't rewrite the F-16's core flight software. It interfaces with the existing controls and sensors, sitting alongside them, switchable. That's a deliberate architectural choice, and it's the harder engineering problem than "make the AI fly well." Anyone who's shipped a model into a system where failure has physical consequences knows this: getting the model to perform is maybe 30% of the work. Getting the handoff right — the sensor fusion for "is the AI still tracking reality," the latency budget for "how fast can a human reclaim control," the failure mode for "what if the switch itself glitches" — that's the other 70%, and it's the part that never makes the headline.

There's also a quieter detail worth pulling out: this ran on an operational-fleet F-16, not a purpose-built experimental aircraft like the X-62A VISTA that came before it. VISTA proved an AI could win a dogfight in a jet built specifically to test that. VENOM proves you can retrofit the autonomy into a jet that already exists, that's already in service, without touching the underlying avionics. That's the difference between a research result and a program that scales. I've seen the same gap kill more "AI pilots the whole thing" projects than any model accuracy number — the demo works on the bespoke rig, and then nobody can afford to retrofit the fleet.

It also lines up with something I've believed for a while: the interesting engineering in high-stakes AI isn't happening in the cloud, and it isn't happening in the model weights alone. It's happening at the interface — where inference meets hardware that was never designed with AI in mind, and where someone has to decide, in advance, exactly how much authority the model gets and exactly how fast that authority can be revoked. Bolt-on autonomy on legacy hardware, with a human-in-the-loop switch, is a much more honest picture of where defence AI is heading than "autonomous fighter jet" implies.

None of this means the milestone isn't real. Handing live flight control to an AI agent mid-sortie, on a combat-capable airframe, is a genuine step — DARPA has been building to this since the X-62A dogfight tests, and getting from "wins a simulated dogfight in an experimental jet" to "flies an operational F-16 with a pilot who can reclaim control instantly" is real progress. I just don't think the AI is the part that should get the credit. The switch is.

If you're building anything where an AI model touches a system that can hurt someone if it's wrong, copy the switch, not the hype. Design the override first. Make it fast, make it unambiguous, and make sure it works even when the model is confidently, silently wrong. The model is the easy 30%.

Takeaway: the milestone that matters in high-stakes AI is never "the model can do the task." It's "we know exactly how to take control back, and it's fast enough to matter."