Pentagon wants AI that turns satellite, radar and video into missile alerts in under five seconds

The Pentagon wants an AI that can turn live satellite, radar and video feeds into a missile-threat alert in under five seconds, and a Sept. 24 solicitation from its Defense Innovation Unit spells out three graded ways humans can sit in, on, or out of that decision loop.

The ask is the "Space Threat Intelligence Synthesis Engine," a piece of software DIU wants contractors to build under solicitation PROJ00716. The system has to fuse "massive streams of multi-source data, including live video, satellite imagery, radar and sensor feeds, geospatial data, and classified intelligence reports" into a single, continuously updated threat picture, per the solicitation language reported by Marine Corps Times.

The latency target is the most concrete spec in the document. No more than five seconds can pass between data arriving and the alert appearing on an operator's screen, and preferably no more than two. Sustained throughput has to hit 20 to 30 megabytes per minute, with bursts up to five gigabytes. The system also has to produce human-readable visualizations for frontline operators and machine-to-machine APIs that can drive automated command-and-control workflows, the two output channels that decide who, if anyone, sees the alert before a response is triggered.

DIU names three workflows the same engine has to run: analyst-in-the-loop, where a human reviews and approves each alert before any downstream action; human-on-the-loop, where the AI generates and acts on alerts while a human supervises and intervenes only on low-confidence cases; and fully automated command-and-control, where the AI's outputs feed directly into other machines with no human step in between. Each mode produces "confidence-scored alerts," per the solicitation, so the same engine can shift from one mode to another depending on the operator, the threat type, or the rules of engagement. DIU is not picking a side. It is asking contractors to build the toggle.

The Pentagon is not just buying a tool. It is asking for code the broader defense-AI community can read, audit, and extend, an unusual posture in a missile-defense procurement, where proprietary stacks are the norm. The likely intent is a wider bidder pool and a faster path to follow-on work than a closed contract would allow.

The backdrop is what gives the solicitation its operational weight. The document explicitly cites the saturation attacks seen in Ukraine and Iran, where cheap drones and ballistic missiles have run up against a "perpetual shortage of expensive interceptors, such as the Patriot air-defense system." When the cost of the response is measured in multimillion-dollar missiles, the constraint moves up the stack: the question is no longer which interceptor to fire, but whether to fire at all, and that question has to be answered faster than a human staff can answer it on a good day. DIU's two-to-five-second window is the time budget the Pentagon is willing to give a machine to answer it.

The public excerpt of the solicitation, as of Sept. 11, leaves three things open: how DIU will grade the trade-off between speed and accuracy, which contract vehicle will carry the award, and what the dollar figure looks like. Contractors have to show "significant speed and accuracy improvements in event detection, characterization, and attribution compared to existing baseline systems," but the rubric for "significant" lives in the unhydrated tail of the source document and in the read-out of the Sept. 24 responses.

The document does, however, put the Pentagon's preferred human-machine boundary on the public record three weeks before any code is fielded. Allied procurement officers, congressional staff, contractor capture teams, and the AI safety community can read the same graded autonomy menu and push back on it before any system is wired into a sensor. The Sept. 24 deadline, in effect, opens a public comment window on a specific design choice, written in procurement language.