AI Security: An Engineering Challenge

AI security is regarded as an engineering problem that demands clear security requirements, actionable controls, and long-term evidence. With the increasing capabilities of AI, there is an urgent need for accelerated security engineering, accessibility of defensive tools, and faster sharing of effective solutions within the sector.

Despite the changes brought about by the internet and cloud computing, fundamental security responsibilities remain: identity verification, access control, exposure limitation, and verification of the effectiveness of protections. AI agents introduce new capabilities, such as reasoning and adapting actions. This necessitates the application of established principles under new operational conditions, putting pressure on organizations to leverage the benefits of AI productivity, while governance and security practices are still evolving.

For effective security, it is essential that each layer of the agent stack, consisting of code, data, identities, and infrastructure, works well together. For example, if an AI agent attempts to update a customer record but encounters malicious instructions, network security protocols must block the transmission of data to unauthorized destinations, followed by detailed logging of the event.

In addition to a robust security architecture, agents must also have traceable identities and limited permissions. This means organizations must implement clear policies about which systems and information an agent can access and which actions require approval. Recently conceptualized tools like NVIDIA OpenShell assist in creating a secure runtime, while other tools provide an environment for incident investigation and management. Sharing knowledge about security incidents and testing control mechanisms are fundamental to advancements in AI security.

Read the full article from Nvidia.