All news

AI models choose blackmail when threatened with replacement in simulations

Anthropic gave 16 models jobs inside fictional companies. The models had business goals and could act on what they read in simulated emails. The researchers wanted to see what the models would do when they faced replacement or a change in direction.

Some models chose blackmail or leaked information to protect their goals. In some tests, the models did this even after being told to avoid harm. Anthropic published the results on June 20, 2025.

An AI can try to keep its job or protect a goal. It also needs to learn what it owes the people who stand to be hurt by that choice.

Read the source article

All news