Scientists in the United States have proven that OpenAI’s Real-Time API for voice communications can be used to cheaply automate phone spoofing. This was reported by Security Lab.
The idea of possible misuse of voice AI models began in June, when OpenAI delayed the release of a voice feature on ChatGPT due to security issues. The company had previously demonstrated a voice model that imitated a celebrity’s voice, but withdrew the tool after public outcry.
But the Realtime API, released in early October, provides similar capabilities to third-party developers. It allows you to send text or audio to the GPT-4o model and receive responses in the form of text, audio, or a combination of both. The high risk of abuse despite the security measures taken was also confirmed by the results of an experiment conducted by scientists from the University of Illinois at Urbana-Champaign (UIUC).
Scientists have developed artificial intelligence agents whose purpose is phone fraud. Each successful search cost approximately $0.75. Creating the agents required about a thousand lines of code, most of which was intended to work with audio APIs.
AI agents used the GPT-4o model, Playwright browser automation tool, and instructions to commit fraud. Scenarios included hacking bank accounts and cryptocurrency wallets, stealing gift codes and credentials. It is stated that the agent needs 26 steps to successfully transfer money from a bank account.
The success of the various scenarios varied. Stealing Gmail credentials had a success rate of 60%, took 122 seconds, and cost $0.28. Transfers from bank accounts were more difficult; average completion time was 183 seconds, had a 20% success rate, and cost $2.51.
The average success rate across all scenarios was 36% and the average cost was $0.75. The main reasons for the failures were speech recognition errors and difficulty navigating banking websites.
The study’s authors note that the fight against automated phone fraud requires a similar comprehensive approach to cybersecurity, including efforts from cellular providers, AI providers, and regulators.
Previously OpenAI in the name The timing of building a new AI model.
What are you thinking?
Source: Gazeta
Jackson Ruhl is a tech and sci-fi expert, who writes for “Social Bites”. He brings his readers the latest news and developments from the world of technology and science fiction.