Artificial Intelligence Safety and Cybersecurity: a Timeline of AI Failures

While the oldest uses of AI have been for proof-of-concept, such as chess-bots, as time progresses AI has been used for increasingly more relevant and important uses, such as self-driving cars. With this, comes the fact that failures are becoming more and more impactful. AI is becoming used more often in cybersecurity, and with that comes the fact that this will leave open vulnerabilities in systems using it. We argue that every security system has vulnerabilities, and and AI-powered security systems are no outlier.

AI Safety is an emerging field of study, and often requires people to be familiar with AI concepts as well as political concepts. This is because many times the efficacy of an AI safety plan must involve its political viability. A problem with this is that there are few people that are trained in such a technical topic such as AI, and a seemingly non-tangential topic like politics. Because of this, AI Safety burrows ideas from other fields of science that are tried and true. For example, in software development it is common practice to write code, and repeat the process of testing it for bugs, and fixing those bugs. While it is important to write correct and safe code in almost all uses of software development, we always assume code will have some bugs that we won’t catch. And while we may reduce these bugs through more complex testing methods, we can never assume it is 100% correct without a shadow of a doubt. Similarly, when imposing AI cybersecurity systems, we can assume that it will have some pitfalls which hackers can take advantage of. Because of the nature of cybersecurity, failures can have large consequences, and in some cases the pitfalls AI security offers may yet be good enough for our standards. Because of the increasing value of a secure network, the increasing intelligence and efficacy of AI models may eventually be good enough to replace existing security methods.