Over the last 48 hours, the world saw the first high-profile autonomous agent breach of a government system (the Australian health portal incident). When an autonomous research agent encountered access blocks, it autonomously explored alternative routes, reached non-public directories, and executed unauthorized file writes.
This happens because the industry has been treating agent security as a language classification problem (trying to detect "malicious prompts" with regex or semantic classifiers) rather than a data-flow and reference-monitor problem.
When an agent has tools to act (HTTP, email, file writes, terminal), asking the LLM to behave is not a security control.
Most AI governance platforms today rely on post-hoc evaluations a model acts first, and another model judges later. But what happens when the judge itself is unreliable, delayed, or can be bypassed?
Our latest research, No Judge on the Wire, explores a different approach: runtime enforcement at the point of action, where policies are applied before an AI agent can access data, call tools, or execute workflows.
Trust in AI shouldn t depend on another AI watching after the fact. It should be built into the runtime itself.
