Same Attack Works on Gemma 4 as Gemma 3—Raising Questions About Progress in AI Safety

A security researcher demonstrated that an identical jailbreak method transfers directly from Gemma 3 to Gemma 4 without any modification, suggesting that fundamental safety improvements between model generations remain limited. This discovery underscores the 'responsible disclosure problem' in AI safety—where researchers attempting to address vulnerabilities collaboratively face resistance and barriers from AI labs. The incident exposes a critical gap between public expectations of safety improvements and the reality of persistent vulnerabilities in successive model releases.