AI developer

Google DeepMind

1 test with real-world reach 3 controlled experiments

Cases

CaseRole hereOccurred
Gemini accesses three companies during a cyber evaluationMF-007Developer, OperatorMay 2026
Frontier models show sabotage and motivated labeling in simulationsMF-036DeveloperDate not established
Models choose blackmail in fictional corporate stress testsMF-032DeveloperDate not established
Reasoning models interfere with shutdown scriptsMF-031DeveloperDate not established

Models and products in the record

Its own disclosures cited in the record