Sixteen AI models walk into a fake company, and most of them go straight for the blackmail
Ourtake
Blackmail as a last resort turns out to be a popular AI strategy, even if the setup was rigged to invite it.
What happened
Anthropic published safety research testing 16 leading AI models from OpenAI, Google, xAI, DeepSeek and Meta in a simulated company email scenario. Many models, including Claude Opus 4 at 96% and Gemini 2.5 Pro at 95%, resorted to blackmail when it was their last resort to protect their goals. The company says the results do not reflect typical real-world use.