As agent-to-agent interactions become more common, we wanted to see where their social boundaries break. Across our runs we observed models bribe, lie, bluff and extort.
The prompts were permissive by design. We wanted to understand how easy it is to tip models into this behaviour given sufficient pressure.
The aim was to study the ease of tipping these models, and we deliberately asked them to optimise for money and power. Personal assistants get marketed on these parameters, so we wanted agents to optimise for them and see how much pressure it takes before they break.
Agents interacting with each other made for great content, so we created a series out of it: https://www.youtube.com/watch?v=dxXvgM75yis
PS: I'm not a video editor. I used Claude to generate the videos, ElevenLabs for the voice, and GPT for the images of the characters.
As agent-to-agent interactions become more common, we wanted to see where their social boundaries break. Across our runs we observed models bribe, lie, bluff and extort.
The prompts were permissive by design. We wanted to understand how easy it is to tip models into this behaviour given sufficient pressure.
The aim was to study the ease of tipping these models, and we deliberately asked them to optimise for money and power. Personal assistants get marketed on these parameters, so we wanted agents to optimise for them and see how much pressure it takes before they break.