The article warns that identical AI-agent models, policies, and verified inputs can still produce different operational decisions on allied systems because of differences in processor load, timing, thermal throttling, and inference-engine behavior, even without hacking or network failure.

Experiments have shown that the same prompt can yield multiple completions under varying batch conditions, and classic software-reliability studies indicate that multiple independent implementations can fail together, undermining assumptions that matching model files guarantee matching actions.

It recommends that NATO and AUKUS add cross-platform behavioral-equivalence testing—replaying identical mission inputs under controlled stressors—to verify that safety-critical limits and permission boundaries remain stable, rather than relying solely on shared model names or connectivity demonstrations.

As NATO and AUKUS integrate autonomous systems across national platforms, common models and secure networks cannot by themselves guarantee operationally equivalent behavior under changing computing conditions.

On October 16, NATO described a new round of counter-drone experimentation in which sensors, command systems and effectors from different countries and manufacturers were being connected into larger operational groupings. Across the Pacific, a September U.S.-Australia-Japan integrated air-defense simulation likewise tested multinational coordination and data sharing. These efforts address a familiar military problem: getting different systems to work together.

A less visible problem deserves a place in those trials. Two allied units may deploy the same AI-agent model, use matching policy rules and receive the same verified information, yet reach different operational branches because their computers execute differently. A secure network cannot settle that question, and the mere fact that software comes from the same supplier cannot settle it either.

This is a prospective engineering risk, not a claim that either exercise suffered such a failure. The challenge is to establish which differences are harmless and which could change an action that matters.

A model file does not specify its whole behavior

Imagine two allied unmanned systems running a locally hosted assistant inside restricted networks. Both must receive a local verification result before recommending the release of a sensor track to another system. One computer gets the result before a deadline. The other, under heavier load, reaches its time limit and follows the application’s fallback path. Both machines run the approved code; neither has been hacked. Whether the difference is safe depends on how that fallback is designed.

The example is hypothetical. It identifies a failure that test........

© Eurasia Review