Smack Wins $7.5m SOCOM Deal for Omega AI Planner

US Special Operations Command (SOCOM) has awarded Smack Technologies a $7.5m prototype agreement for Omega, the Austin start-up’s AI planning tool for military commanders, the company announced on 26 October. The deal is an Other Transaction (OT) agreement, a flexible contracting route used for prototypes outside standard procurement rules.

Smack also posted the award on X. SOCOM has not published its own notice, and TSN found none on the command’s site, SAM.gov or the Defense Department’s contract announcements. The amount and terms therefore rest on Smack’s account. DefenseScoop, which interviewed chief executive Andy Markoff before the announcement, described the agreement as “worth up to $7.5 million”, suggesting the figure is a ceiling rather than money already paid.

In brief

Smack says SOCOM has awarded it a $7.5m prototype OT agreement for Omega, its planning and decision-support software.
Smack says Omega runs on small models that can work without a constant cloud link, and that it vets new data before any model learns from it.
No SOCOM or Pentagon notice has been published, so the details come from the company.

Omega builds and repairs plans for commanders

Smack describes Omega as “Smack’s AI decision platform for commanders”. In an explainer published in August, the company said the software takes in planning documents, maps and other inputs, interprets the mission and produces courses of action, with the risks of each set out for the person making the decision.

Markoff gave DefenseScoop a plainer summary. “Commanders can generate operational objectives, the tactics that achieve them, and the tasks that achieve those — all linked as a single plan,” he said in an email. “As the fight changes and assessments come in, Omega enables commanders to repair the plan in seconds.”

Smack says the award “is part of SOCOM’s initiative to accelerate the application of frontier technologies into deployable warfighting capabilities across domains”. Markoff, a former Marine Raider, argued in the release that fusing large volumes of mixed data for decisions “is a capability that does not exist today”.

The company was co-founded by Marine Corps special operations veterans, according to DefenseScoop, and has earlier prototyping contracts with the Marines, Air Force and Navy. Markoff told the outlet he expects Omega to move into production across several US military branches in the coming months. That is his forecast, not an announced order.

SOCOM has been taking on a larger AI and autonomy role. Earlier this month defence officials said the Pentagon’s autonomous warfare group was being folded into a new task force inside SOCOM. Smack is also one of several start-ups chasing the same planning problem: Watchtower Labs came out of stealth last week with $35m to build AI decision tools for military logistics.

Offline learning and small models are Smack’s answer to poisoned data

The most useful part of Markoff’s account is about failure, not speed. He told DefenseScoop that Omega was designed to withstand adversarial AI attacks, spoofed sensor data and deliberate data poisoning in three ways. First, “Omega doesn’t depend on constant cloud connectivity”: the models are small enough to run on local servers at separate sites, each working within limits set by higher headquarters, solving problems locally when communications drop and sending data when a link returns. Second, it pairs classical planning algorithms, which are deterministic and explainable, with generative methods. Third, it learns offline. “Adversaries will flood the battlespace with spoofed and poisoned data, so we validate new data before any model learns from it,” Markoff said. “Because we run many small models, we can retrain only the model that needs it.”

Those are sensible design goals, and they address real risks. A planning tool that learns from whatever data arrives can be steered by whoever controls that data. Keeping a deterministic, explainable layer also gives staff a way to check why a plan was proposed. Smack says that where a third-party large language model is used, its role is “narrowly scoped” and grounded in the company’s own context graph.

None of it has been independently tested in public. Smack has not published evaluation results, and SOCOM has made no public statement about the agreement. The questions that matter are the ones a prototype phase should answer: how the vetting catches poisoned inputs, how often plans need human correction, and who signs off. TSN has reported on Army Cyber Command’s interest in AI agents that act with minimal human oversight, and on an Air Force test in which a pilot directed a combat drone in flight. In each case the record on human control matters more than the speed claims.

The first thing to look for is a SOCOM statement setting out what the prototype is meant to prove and how it will be judged.