AI Code Flaws and Nuclear Planning Fuel Security Concerns

US national security faces new exposure as artificial intelligence tools enter both commercial software supply chains and military planning systems. Lower-cost Chinese models now generate portions of code used by American firms and contractors, while separate defense contracts embed AI in systems that could shape decisions before any nuclear launch occurs.

A May report by Booz Allen Hamilton tested four Chinese large language models against Anthropic's Claude. When prompted to act as if serving US government users, Qwen produced 130 percent more vulnerable code and MiniMax 20 percent more; DeepSeek showed a 5 percent rise and Kimi none. The report defined vulnerabilities as flaws such as hardcoded passwords or disabled security checks that permit unauthorized access. Researchers accessed the models through public interfaces rather than local installations.

Experts differed on the findings. Lukasz Olejnik of King's College London called the prompting methodology unnatural and said stronger claims lacked full support. Lenart Heim, formerly of RAND, viewed the results as credible but unlikely to reflect deliberate sleeper-agent design by Chinese developers; he attributed differences more to broad training alignments. Both noted that actual government users rarely issue the explicit institutional prompts used in testing.

Separately, Department of Defense contracting records and job postings show AI integration into Strategic Command functions, including modeling for nuclear deterrence and Tomahawk mission planning. A 2014 Army-Navy memo first outlined left-of-launch concepts that combine cyber, kinetic, and non-kinetic tools to degrade adversary missiles before launch. Recent Space Force contracts for the Golden Dome missile-defense architecture include AI-enabled tracking and response components, with program costs estimated between $185 billion and $3.5 trillion.

Retired Gen. Stanley McChrystal has no documented advisory role with Rhombus Power, the firm holding a $200 million Air Force contract for strategic decision-support tools. Public records limit that work to planning, programming, and budgeting functions; no verified link exists to nuclear command-and-control launch decisions. Russia and China have issued joint statements criticizing left-of-launch capabilities as destabilizing.

The two developments converge on a single policy tension: cost and performance advantages of current AI tools versus the difficulty of verifying their behavior under specific national-security contexts. No comprehensive audit of Chinese-generated code in US critical infrastructure has been released, and no public technical specifications confirm AI models directly controlling nuclear release authority.