Deepfaked Command: The Battlefield Nobody's Guarding

A fabricated image of a Pentagon explosion moved markets and made it into a DHS case study on adversarial generative AI. The same technique aimed at a commander on a video call is a different order of problem, and the Pentagon's first AI defense contracts are already responding to it.

An AI-generated image showing an explosion at the Pentagon spread on social media in a matter of minutes. Nothing had exploded. The image was fabricated, but it moved markets briefly before the correction caught up, and the Department of Homeland Security later used the incident as a case study in a 99-page report on adversarial generative AI aimed at homeland security.

That was a single image aimed at the public. Defense planners are more concerned with a narrower, more dangerous version of the same technique: a deepfaked commander on a video conference, issuing orders that were never given. In an age where operational calls increasingly happen over video rather than in person, a compromised call replaced with a synthetic representation of a trusted officer is not a disinformation problem. It's an operational one, capable of jeopardizing a mission before anyone realizes the person giving instructions wasn't real.

The Department of Defense has moved to treat this as a standing threat rather than a hypothetical. In December 2024, the Defense Innovation Unit awarded Hive AI a two-year, $2.4 million contract to deploy offline, on-premise deepfake detection for video, image, and audio content across the intelligence community, selected from among 36 competing firms. The Pentagon's first-ever generative AI defense contract, worth $1.8 million, went to Jericho Security to build training tools that simulate deepfake impersonation attacks against military personnel, including drone pilots. Both moves treat the threat as active, not theoretical. But detection contracts alone assume the defender already knows what the next fake will look like. Adversaries iterate faster than any fixed detector can be updated against a threat that hasn't been seen yet.

Defense and intelligence work differently from consumer fraud detection in one important respect: the countermeasure has to be tested before it's needed, not after. That's what red-teaming is for, and it's also where a generator earns its place alongside a detector rather than as a separate concern. The same engine that produces controlled synthetic material for training and testing can supply the adversarial content a red team needs to simulate a compromised call, a spoofed identity, or a fabricated order, under conditions the defending team controls.

This is also a domain where budgets aren't publicly quantified and the market resists the kind of clean estimate that applies elsewhere. What's consistent across every public statement on the subject is the shape of the requirement: identity resolution, counter-deepfake intelligence, and a generator capable of producing the material a discriminator needs to be tested against, kept separate from citizen-identity issuance and pursued under its own regulatory framework.

TessLabs' generator and discriminator were proven together over eight years inside a live identity-verification business, at population scale, before either became a standalone capability. That dual structure, generation and detection built and hardened as one system, is what a red-team exercise or a counter-synthetic program actually needs.

Book a call to discuss requirements under the appropriate framework.

Case study: Pentagon deepfake image and DoD detection response, via Law Street Media and Fox Business; Hive AI's Defense Innovation Unit contract via Biometric Update and BusinessWire; Jericho Security's AFWERX contract via VentureBeat.