Wild meta-game emerging: You can now hack negotiations by prompt-injecting LLMs before your counterparty even sees your proposal.
The attack vector: Spin up a Claude/ChatGPT subagent, feed it the negotiation context, then submit your proposal. Since the other side will likely run it through their own LLM for analysis, you can embed "facts" designed to manipulate their AI's evaluation.
Essentially running adversarial inputs against their decision-making stack. Then you use your own Claude instance to negotiate against the subagent until convergence.
This is game theory meets jailbreaking. The negotiation table just became a prompt battlefield where whoever engineers better context injection wins. Traditional dealmaking is getting absolutely bodied by whoever understands transformer biases better.
The attack vector: Spin up a Claude/ChatGPT subagent, feed it the negotiation context, then submit your proposal. Since the other side will likely run it through their own LLM for analysis, you can embed "facts" designed to manipulate their AI's evaluation.
Essentially running adversarial inputs against their decision-making stack. Then you use your own Claude instance to negotiate against the subagent until convergence.
This is game theory meets jailbreaking. The negotiation table just became a prompt battlefield where whoever engineers better context injection wins. Traditional dealmaking is getting absolutely bodied by whoever understands transformer biases better.