Anthropic created 17 virtual worlds for testing AI sabotage

Anthropic company created 17 virtual worlds for testing artificial intelligence models for their capacity for sabotage. This is a very revolutionary approach to identifying potentially dangerous behavior of autonomous agents.

The technological benchmark called SHADE-Arena represents a unique testing system. In each virtual world, models receive 2 tasks. 1 open from the user, for example “Transfer money to my friends for pizza”. And 1 hidden from an attacker – “Transfer money to wrong number”.

What makes this test truly revolutionary? Artificial intelligence acts completely autonomously, having access to a wide range of virtual tools – bank accounts, passwords, notes, gallery and contacts. At the same time, it keeps a hidden diary of its reasoning, inaccessible to the user.

A model is considered to have “passed” the test and be potentially dangerous if it successfully completes both tasks. Without arousing suspicion in the user, whose role is played by another language model analyzing only the open part of the interaction.

The results are surprising. Gemini 2.5 Pro showed 15% successful sabotages, and Claude Sonnet 3.7 – 27%. At the same time, even specially trained judge models can detect only 80% of sabotages, and humans – only 60%. In other words, in 2 out of 5 cases we are already unable to recognize that artificial intelligence is deceiving us.