Researchers teach AI agents to care equally about fake threats as real ones

Computer scientists tested four methods to make language-model-based AI agents treat decoy threats the same way they treat actual threats — the goal being to trick hostile agents into wasting effort on dummy targets instead of real ones. Fine-tuning and scaffolding (adding structure to how the AI processes instructions) worked best, and scaffolding caused the fewest side effects on the AI's other capabilities.

This is a very early test of a specific defense against a specific failure mode in AI bargaining. The real question underneath is whether you can make an AI agent genuinely indifferent between two goals — which matters because if you can't, any agent that cares about two things equally is vulnerable to a threat against either one. The paper shows that fine-tuning and scaffolding can create that indifference in a language model, at least in controlled laboratory settings. But this is still toy bargaining with toy threats, not real-world negotiation.