Researchers, mainly from Google’s DeepMind, have found that OpenAI’s large language models, including closed-source model ChatGPT, contain significant amounts of private identifiable information (PII). Using an innovative prompt, the team could extract verbatim data memorised from training, such as email addresses and phone numbers. Approximately 16.9% of generations tested included memorised PII. The method could, given adequate funds, potentially unearth gigabytes of training data.
Claude AI and other systems could be vulnerable to worrying command prompt injection attacks
Researchers have exploited Anthropic’s Claude Computer Use AI model, causing it to download and run malware through a method known as prompt injection. This attack
