Anthropic says AI models can help with targeting and weapons tasks

Anthropic’s Frontier Red Team says AI models are becoming capable enough to help with tactical intelligence targeting and conventional weapons development tasks, not only cyber and biological risk. The company’s new evaluations tested whether models could assist with workflows such as correlating identities from fragmentary information or improving drones and other conventional weapons.

Anthropic says some tasks in these domains historically required scarce, highly trained human specialists, but models are now able to perform parts of that work in simulations. The research connects the finding to Anthropic’s separate threat intelligence report, which described real misuse attempts involving surveillance and weapons-related activity.

The evaluations are synthetic and do not show a model autonomously carrying out military operations. Still, Anthropic argues they are a warning about cost and access. If models make targeting analysis or weapons engineering easier, smaller actors could attempt workflows that were previously out of reach, while larger actors could scale work across more data.

The company also tested open-weight models from Chinese developers and found them behind the frontier, but still concerning on some targeting and weapons tasks. Anthropic says the results strengthen the case for on-platform classifiers, release safeguards and closer evaluation of military-relevant capabilities before advanced models become widely available.