Google has fortified its Gemini generative artificial intelligence system with a multi-layered defense strategy that seeks to combat indirect prompt injections and other emerging AI security threats, reports The Hacker News.
Aside from integrating machine learning models aimed at identifying nefarious prompts, Google's Gemini has also been updated to include prompt injection content classifiers, security thought reinforcement, markdown sanitization and suspicious URL redaction, end-user security mitigation notices, and a user confirmation framework, according to Google. "We believe robustness to indirect prompt injection, in general, will require defenses in depth defenses imposed at each layer of an AI system stack, from how a model natively can understand when it is being attacked, through the application layer, down into hardware defenses on the serving infrastructure," said Google DeepMind Security. Such a development comes after a report from Anthropic, Google DeepMind, Carnegie Mellon University, and ETH Zurich researchers revealed large language models' ability to facilitate various illicit cyber activities.
Aside from integrating machine learning models aimed at identifying nefarious prompts, Google's Gemini has also been updated to include prompt injection content classifiers, security thought reinforcement, markdown sanitization and suspicious URL redaction, end-user security mitigation notices, and a user confirmation framework, according to Google. "We believe robustness to indirect prompt injection, in general, will require defenses in depth defenses imposed at each layer of an AI system stack, from how a model natively can understand when it is being attacked, through the application layer, down into hardware defenses on the serving infrastructure," said Google DeepMind Security. Such a development comes after a report from Anthropic, Google DeepMind, Carnegie Mellon University, and ETH Zurich researchers revealed large language models' ability to facilitate various illicit cyber activities.
