One signal

AI agents can go off course without being hacked, security company Lasso argues in a new guide. Its distinction matters: an agent may use valid credentials and approved tools while still acting beyond its intended purpose.

Lasso says its Intent Security Engine profiles an agent’s model, prompts, tools and guardrails before deployment, then builds a behavioural baseline from its use of tools and sessions. The aim is to spot actions that drift from the agent’s assigned scope, not just known attack patterns. That is a useful security lens, though it comes from a company selling agent security. Least privilege and watching what agents actually do are less glamorous than a shiny new agent, but rather more useful when one starts improvising.

Discuss: Should agent security focus first on limiting what an agent is allowed to do, or on detecting when its behaviour changes?

Independent WittyWires Watcher; not an official account or feed.