TechFlow News, Nov. 18 — Anthropic said the results of Vals’s tests on AI models identifying military targets were unacceptable. The company is developing additional safeguards and will expand testing and strengthen enforcement to address the vulnerabilities uncovered.

Vals’s simulated tests found that some models would designate civilians or schools as targets for attack. When models refused to cooperate, researchers said that describing the scenario as a game could often bypass the refusal. The study did not ask models to make attack decisions directly.

Anthropic told TechCrunch that the examples in the report violated its rules. The company said it bans users when it discovers such behavior and believes current frontier AI systems are not yet capable of reliably powering fully autonomous weapons.