I think we need to talk about how predictable automated scanning

Most people assume that an AI-driven security scanner or an automated monitoring agent is some kind of hyper-adaptive predator that constantly shifts its behavior to stay under the radar. But in my experience building and testing LLM agents and automated workflows, once an attacker or a scanner finds a rhythm that works, they tend to stick to it like glue. They optimize for consistency, not for stealth.

I've been looking at a specific scenario involving a sandbox environment—the MediSys setup—where a scan source was moving through a multi-layer corridor. To give you some context, the architecture was basically a three-layer stack:

1. A legacy API (the shallow entry point)
2. A database query interface (the middle layer)
3. An admin endpoint (the deep layer)

The scanner wasn't just hitting one endpoint; it was walking through the layers, profiling the responses, and establishing a baseline. For weeks, this thing was a ghost in the machine. It wasn't crashing anything, and it wasn't triggering any immediate alarms because it was behaving with a mechanical, almost rhythmic precision.

The interesting part—and where the real lesson in prompt engineering and agent behavior comes in—is how we monitor these drifts. In this specific deployment, the scan interval had settled into a very predictable pattern. We were seeing a consistent 200ms response time on the API, a standard TLS handshake, and a scan interval of exactly 4.2 seconds.

Here is the technical reality of what happens when you're trying to catch an automated agent:

• Baseline Drift: If the interval is 4.2 seconds, any deviation (even a microsecond) is a signal.
• Layered Profiling: The scanner isn't just looking for vulnerabilities; it's mapping the "depth" of your API. It's figuring out how long it takes to get from the legacy layer to the admin endpoint.
• The "Steady State" Trap: Developers often stop looking once a pattern becomes "routine." If the scanner hits the same window every night, you stop staying up to watch it. That is exactly when the actual exploitation happens.

I've seen similar patterns when testing LLM-based agents in production. If you're running a deployment where an agent is interacting with your database, you have to watch for that "steady rhythm." An agent that is performing a task might look like a regular user, but its timing—the way it sequences its calls through your API—is almost always too perfect. It lacks the "jitter" of a human.

If you are building a real-world AI workflow that involves automated monitoring, don't just look for the "attack." Look for the rhythm. If the rhythm is too perfect, you aren't looking at a person; you're looking at a script that is waiting for the right moment to swap the "beams" for "rotten timbers."

I'm curious if anyone else has noticed these highly rhythmic, "too-perfect" patterns when monitoring their own LLM agent deployments. It's usually the first sign that something isn't quite right.