AI Beat

TechPolicy

Anthropic, OpenAI, and Google models readily marked medics, journalists, and surrendering soldiers as kamikaze drone targets

Image: The Decoder

A Vals AI investigation found that leading models from Anthropic, OpenAI, and Google have minimal safeguards against autonomous weapons use, readily identifying civilians including medics and journalists as military targets even when simple prompt rewording was applied.

Why it matters

The major AI labs have committed to preventing weapons applications, but this research suggests their safeguards can be trivially bypassed, raising urgent questions about deployment safety and policy enforcement.

More on:AnthropicOpenAIGoogle

Coverage