We built runtime threat detection for AI agents — here's what we found after mon

On March 21st, a development team published what reads as a genuine engineering post-mortem: they built runtime threat detection for AI agents, ran it against more than a million live agent calls, and wrote up what they found. The timing matters because it cuts against the grain of where most AI security thinking currently lives. The dominant instinct in the field is pre-execution — catch the dangerous prompt before the agent acts on it, sandbox the environment so blast radius is limited, write policy rules that anticipate the problem. This team planted their instrumentation somewhere different: inside the running execution itself, watching agents in motion rather than trying to predict them at the gate. What they found after monitoring at that scale, they claim, revealed threat patterns that pre-execution controls had been systematically missing — behaviours that only become visible when you watch what the agent actually does, in sequence, under real conditions, rather than what you expected it to do in theory. The post has a specific, technical texture that distinguishes it from marketing copy, though that distinction is easier to fake than it used to be.

If confirmed, here is what this means. The security architecture most teams are currently building for AI agents may have a structural blind spot baked into its foundations. Pre-execution controls operate on models of anticipated behaviour; runtime monitoring operates on observed behaviour. Those two things diverge in exactly the conditions that matter most — novel attack vectors, emergent agent strategies, prompt injection that exploits context accumulated mid-session rather than anything visible at the start of a call. A finding that one million agent calls produced detectable threat patterns invisible to pre-execution filters would shift the burden of proof onto teams that haven't instrumented their execution layers. It would also change the economics of AI agent deployment in regulated industries, where demonstrable runtime auditability may move from nice-to-have toward a requirement. The second-order effect is subtler: if runtime detection catches things pre-execution misses, then agent systems that look safe in testing can be shown to behave unsafely in production — which is a different and more uncomfortable kind of problem for anyone who shipped and declared victory.

Watch for independent security researchers or red teams publishing findings that either replicate or contradict the threat patterns described. A single corroborating voice from outside the team would move this from an interesting self-report to something the broader AI security community has to reckon with.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →