Hi Thomson, thank you for the question - there is a wide array of vulnerabilities out there but they can often be grouped into themes of the attack someone is using. At the moment, what we are seeing is that most (~90%) attacks exploit the capabilities of an AI system rather than the model itself - this could be tactics like impersonation, scaling and amplification, or "sockpuppeting" (hiding behind a false identity to deceive). There is a great Google DeepMind paper on Generative AI Misuse (Taxonomy of Tactics) that is a really interesting read to understand the threat landscape :) hope that helps but happy to answer any follow-up questions!

26 September, 2024 12:32

Thanks Tom

26 September, 2024 12:32

there are many - here are a few headlines: data poisoning – data leakage – ai supply chain attack – inference attack – extractioj – prompt injection – insecure output validation – model denial of service

26 September, 2024 12:35

Thanks Rick

26 September, 2024 12:35

extractioj -> extraction (typo)

26 September, 2024 12:35

Hi Thomson, here is the link to that paper from Google DeepMind that I mentioned: https://arxiv.org/abs/2406.13843

26 September, 2024 12:49

Thank you Tom

26 September, 2024 12:50