Lately I’ve been thinking about time.

Uptime, pulse checks, execution time, response time. We’ve always treated these as health metrics. They tell us whether a system is alive, responsive, and performing as expected. But what if they’re also security metrics?

That idea isn’t entirely new. At the network layer, covert timing channels, beaconing detection, and behavioral baselining have shown us for decades that the intervals between events matter. Attackers have long understood that rhythm carries information. More recently, researchers have demonstrated timing side-channel attacks against LLMs, using cache latency to infer private prompts and token cadence to fingerprint model outputs.

What I find interesting is the imbalance. Most of the research asks, “How can timing be exploited?” Very little asks, “How can timing help us defend?”

A 2026 systematic survey of LLM-agent security identifies temporal anomaly detection infrastructure as an open research gap, noting that current agent deployment frameworks don’t even support the behavioral baselines such an approach would require. Even then, the discussion largely focuses on session-level behavior. The rhythm within a single execution, the space between observable events, remains largely unexplored.

Maybe time isn’t just metadata, maybe it’s another dimension of observability that we’ve been overlooking.

Time tells you duration and speed. But read carefully, it also reveals location, choke points, and absences, the things that didn’t happen when they should have.

I’ve started exploring this in my own observability work, measuring behavioral changes & entropy across inter-arrival intervals and treating rhythm as signal rather than noise to smooth away.

Curious to know who else is working on the defensive side of temporal behavior, especially for agentic systems or any thoughts or opinions on this topic.

Reference: “A Systematic Survey of Security Threats and Defenses in LLM-Based AI Agents: A Layered Attack Surface Framework,” arXiv:2604.23338 (2026). https://arxiv.org/abs/2604.23338