Startups use AI to create military software prototypes

The creation of complex software is no longer the exclusive preserve of programming specialists. A project led by the United States Department of the Air Force, in partnership with the MIT Lincoln Laboratory, demonstrated that military personnel without any technical training can develop functional tools using only language models.

The experiment, part of the Phantom program, sought to understand whether artificial intelligence Generative technology could reduce the time and costs of the traditional strategic systems development cycle. The central idea is to allow professionals who master battlefield problems to translate their needs into viable digital solutions.

The concept of vibe-coding

Researcher Joshua Lynch adopted a methodology that he himself dubbed vibe-coding. In this process, the user does not write lines of code manually, but uses prompts in chatbots such as Claude, ChatGPT and Gemini to guide the AI ​​in building and refining logical instructions.

Over the course of three months, the initial objective was ambitious: to create tactical support software for target recognition and communications management. The project showed that, although AI is a powerful assistance tool, it requires a significant learning curve for the user to know how to deal with its technical limitations.

Limitations and scope readjustment

During development, Cadet faced common challenges in language models, such as a lack of hierarchical focus and the tendency of AI to change pieces of code that should remain intact. It was necessary to learn to break problems down into smaller tasks and maintain strict control over the system's logical sequence.

Due to these restrictions, the software prototype original was adjusted for a document processing tool. The end result, focused on the analysis of tactical maps and the generation of mission plans, confirmed that the technology works better as a design assistant than as a ready-to-use solution.

Security and the role of human collaboration

A critical point identified by researchers involves information security. The system even sent sensitive data for external analysis in the cloud without the user noticing, highlighting the risk of using generative models without due technical auditing or knowledge about how the code processes information.

Researcher Laura Niss, who supervised the project, reinforces that AI acts as a communicative bridge, but does not replace the human expert. THE code development AI-assisted software still requires thorough review, especially when it comes to applications that handle sensitive data or high-risk operations.

Conclusion and perspectives

The experiment validated that military personnel can, in fact, prototype useful ideas, but highlighted the gulf that still separates rapid prototyping from the delivery of secure and stable software. Collaboration between experts from different areas remains the fundamental pillar for transforming AI suggestions into effective solutions to real problems.

FAQ

• What is vibe-coding? It is a development method where the user guides an AI via prompts to generate code, without having to write the syntax manually.
• Can AI replace programmers in military projects? No. The study indicated that it functions as a design and prototyping assistant, requiring human validation for security.
• What are the main risks encountered? Lack of code verification and inadvertently sending sensitive documents for processing to external third-party servers.
• Did the project result in an operational tool? No, it resulted in a working prototype for document analysis that demonstrated the viability of the concept, but is not ready for combat use.

Also read: MIT AI creates virtual worlds to train robots.

Source and methodology

This article was prepared based on information published by news.mit.edu, on July 7, 2026. See the original post: How novice coders can develop AI programs for military applications. HTechBD reorganized and contextualized the data for the Brazilian public, without reproducing the source text.

Image: Egor Komarov no Pexels.