Blue41 co-founder Tim Van Hamme spoke with Dominique Deckmyn of De Standaard about how to secure AI agents, following NVIDIA’s launch of its Open Agent Safety Platform.
A central part of the discussion was the difference between placing controls around an agent and monitoring how that agent actually behaves.
As Tim explained, an AI agent combines an AI model with the tools and systems it is allowed to use. Rules between the two can prevent specific actions—for example, blocking transfers above €100. But preventive controls cannot anticipate every sequence of actions an agent may take.
So you also need to monitor whether an agent suddenly makes hundreds of €99 transfers.
That distinction is central to how we approach AI agent security at Blue41. Preventive controls remain important, but they are only one part of the security model. Runtime monitoring provides an additional layer by continuously observing agent behavior and detecting patterns that may indicate manipulation, misuse or unexpected actions.
Recent incidents show why this matters. This summer, an OpenAI agent operating in a test environment unexpectedly accessed systems at Hugging Face, while the behavior remained undetected for an extended period. The incident illustrated a broader challenge: advanced autonomous agents can behave in ways their creators did not anticipate, and those behaviors can be difficult to detect without dedicated runtime visibility.
We explored the detection gap and its implications for securing autonomous agents in our analysis of the OpenAI–Hugging Face incident.
Read the full interview in De Standaard.
If you’re working through how to secure AI agents in production, get in touch.