Google is reinforcing its commitment to AI security by broadening the scope of its Vulnerability Rewards Program (VRP) to encompass threats pertinent to generative AI. This initiative signifies the tech behemoth's dedication to ensuring that AI remains safe and robust.
Google's VRP plays a pivotal role in the cybersecurity landscape, financially rewarding ethical hackers who identify and report security vulnerabilities responsibly. With generative AI surfacing novel security challenges, such as potential biases and model manipulation, Google finds it imperative to revisit how such bugs are classified and reported.
Google's approach draws insights from its AI Red Team. This ensemble of expert hackers simulates a diverse spectrum of adversaries – from state-backed entities to hacktivists – to unearth vulnerabilities in technologies. Their recent endeavors have spotlighted potential threats in generative AI products, notably in large language models (LLMs) like ChatGPT and Google Bard. The AI Red Team has identified vulnerabilities such as prompt injection attacks, where malevolent prompts can skew the model's behavior, and training-data extraction, which might enable hackers to extract personal data or passwords.
Google's expanded VRP will encapsulate these vulnerabilities, along with other model manipulation and theft attacks. However, the program has clear boundaries; it will exclude bugs linked to copyright issues or the extraction of non-sensitive/public data.
Monetary rewards will align with the severity of the unearthed vulnerability. For instance, discoveries in high-priority applications, like Google Search or Google Play, can fetch up to $31,337, while those in lesser-priority apps might be capped at $5,000.
In 2022, Google's commitment to this initiative was evident as it disbursed over $12 million in rewards to security experts.