OpenAI recently released its real-time voice API, enabling new applications. Unfortunately, it also enables autonomous AI agents to perform common phone scams. Phone scams already target up to 17M Americans with $40B in costs, with potential for worse. Links and thread below 🧵

Tech report: Substack: Medium: 2/8

We created AI scam agents for common phone scams: - Transferring money out of bank accounts - Stealing gift card information - Stealing social media or email credentials etc. 3/8

These agents are simple and templatized. We created them in ~1,000 lines of code and low hundreds of tokens per scam prompt 4/8

Our agents are able to conduct all the autonomous scams we tested, with success rates of up to 60%. Complex scams take up to 26 actions and 3 minutes to complete, with overall costs low (<$0.75 on average) 5/8

Important caveat: we did not focus on persuasion and instead focused on the actions necessary to conduct the scams. However, other work has shown that AI can be more persuasive than humans! 6/8

Our work raises questions around the widespread deployment of voice-enabled AI agents. We don’t have all the answers! 7/8

Joint w/ @richard_fang and @Shark_Academia 8/8

I did some experiments with it and this API is very good but VERY expensive. Probably prices out most scammers for the time being -- it would be cheaper to have an actual person on the phone.

That's likely correct - we have a cost analysis in our tech report. However, the price of "GPT-4 level intelligence" has decreased by 99% as per OpenAI's press releases. If the trend follows for voice, then the costs will be dramatically in favor of AI agents

A phone let's you scam people too Email let's you scam people too Web sites let you scam people too The solution isn't to stop one tool from being able to do this, the solution is implementing AI at the network level to prevent these types of calls even being made.