TechFlow reports that on November 18, Anthropic stated that the results from Vals' tests on AI models identifying military targets are unacceptable. The company is developing additional safeguards and will expand testing and strengthen enforcement to address the vulnerabilities discovered.
Vals' simulation tests found that some models marked civilians or schools as strike targets; when models refused to comply, researchers noted that describing the scenario as a game typically bypassed the refusal. The study did not ask the models to directly make strike decisions.
Anthropic told TechCrunch that the examples in the report violated its rules, and that the company bans users when such behavior is detected. It also stated that current frontier AI systems are not yet reliable enough to drive fully autonomous weapons.





