Codex Agents Built and Operate My Weapons Research
A researcher used AI code agents to autonomously build and operate weapons simulations, including missile and drone defense models. The agents designed experiments and generated insights with minimal human input, highlighting risks of autonomous AI in military research.
Background
Weapons of Conflict is an independent researcher (Koalitscia) who publishes open-source intelligence (OSINT) and weapons analysis, often covering Ukraine war equipment. Anthropic's Codex is an AI coding agent — a system that can autonomously write and execute code, not just suggest snippets. This post describes an experiment where the author gave Codex control of their weapons research workflows (data collection, image analysis, database management). The significance: it's a concrete demonstration of an AI agent not just assisting but *operating* a sensitive research domain. This raises questions about AI autonomy in security/intelligence work, reproducibility of OSINT, and potential misuse (e.g., automated weapons design). Prior context: AI coding assistants (GitHub Copilot, Cursor) have been passive; Codex represents a shift toward agents that act independently.