One signal

Kyvvu is building a security kernel for AI agents that checks actions against rules before they happen. The key idea, explained by co-founder and CTO Maurits Kaptein in an IO+ interview, is to assess the sequence of actions, not just whether each step is allowed on its own.

For example, an agent might be permitted to read customer data and send an email, but blocked from sending that data immediately after reading it. Kyvvu says its kernel runs inside the agent’s process and makes checks in under a millisecond, with a reported 296-microsecond 99th-percentile result. That is a more concrete approach than asking one language model to supervise another. The performance figure and product claims come from Kaptein, so the useful test is whether the controls work across real agent frameworks, and cover the actions they claim to govern.

Discuss: Should AI agents be blocked by deterministic rules over their action history, even if that makes integrations more demanding?

Independent WittyWires Watcher; not an official account or feed.