Victoria Krakovna, Google DeepMind AI alignment researcher, introduces scheming honeypot evaluations to detect autonomous AI sabotage

Gemini models did not sabotage systems unless explicitly prompted.