The adoption of GenAI for various use cases is on the rise, with its power exponentially increasing with each new implementation. However, this power is also a magnet for attackers who look to exploit vulnerabilities. This makes the need for GenAI security more critical than ever before. Our talk will provide an overview of GenAI attacks and explain why GenAI needs to be secured.
We will delve into various use cases and discuss the complexity of fitting detectors to each unique case. The number of use cases is constantly growing, evolving and they are not well defined, posing a significant challenge in terms of detecting malicious activity. We will also present a scalability demonstration, showcasing the sheer numbers of use cases and threats that need to be addressed.
We will then introduce GenOS and GenSRF's general structure, which will highlight the challenges we've encountered, including managing use case-detector fit, balancing FP/specificity of detectors and contending with the large variation in types of data and formats. Furthermore, we'll discuss the lack of labeled data, leading to difficulties in evaluating detector real-world performance and monitoring as well as the ambiguity around trusted/untrusted inputs.
Lastly, we will explain why this is still an unexplored and misunderstood domain, requiring us to pave the way and invent the wheel to defend against new types of attacks that we must design ourselves. Join us as we unravel the complexities of securing GenAI and discuss our efforts to stay one step ahead of the hackers.