Text: Chris Cochran, Field CISO & Vice President of AI Security at SANS Institute
The first step? Visibility. Organizations need to understand where AI is being applied within the company. Not just the tools approved by management, but also the AI that appears in workflows, plug-ins, agents, and third-party platforms. Keeping track of a so-called AI inventory will soon be a basic requirement - if it isn't already.
Next, third-party risk is of great importance. An 'AI Bill of Materials' (AIBOM) helps organizations understand which models and data sources are under the hood, and where external risks may arise.
As AI agents become more autonomous, we also need to adjust our approach to security. Agents should be seen as operators on the network, not as traditional service accounts. This means that agents need to have an identity, even if it is temporary in nature. In many cases, temporary authentication and authorization are even preferred. Technologies (think of SPIFFE) can help make such identities possible at scale.
Zero Trust principles still apply. Authenticate explicitly and grant minimal rights. Assume a breach. Continuously monitor the behavior of agents and use segmentation. If something goes wrong, the impact remains limited.
AI on Safer Internet Day
Finally: keep identity and access graphs sharp in focus. Understand where your people have access, where your agents have access, and where those paths intersect. Most AI-related breaches do not stem from the model itself, but from overly broad rights and invisible access paths.
AI can be a huge game changer for companies. The organizations that come out on top are those that combine innovation with discipline and build security into the way AI actually operates on a daily basis.
In the meantime: Happy Safer Internet Day on this 10th of February!
