Most discussions about AI security focus on what models might do wrong. The more urgent issue is what detection systems still cannot see-and side-channel attacks are making that gap visible.
Side-channel attacks target physical factors like power consumption, electromagnetic emissions and processing time, not software code. They can exfiltrate cryptographic keys by measuring hardware emissions. Recent research shows an outside observer can infer the topic of an AI interaction simply by analyzing encrypted traffic patterns. No decryption required.
For two decades, detection has been defined by rules: signatures, thresholds, known patterns. Side-channel attacks don't provide matchable signals. An attacker operating through encrypted channels or AI-assisted workflows can move without triggering a rule.
The practical consequence is straightforward: attackers operate within an environment, and security teams receive no alert. As organizations expand AI use, the proportion of activity falling into this gap increases.
Closing this gap requires detection based on behavioral sequences, not isolated events. The signals already exist in timing, sequencing, and interaction patterns. Deep learning approaches that make side-channel attacks effective can also identify traffic patterns revealing those attacks.
Evan Powell, CEO of DeepTempo, argues security leaders must distinguish between systems making existing detection more efficient and those expanding what detection can observe. That distinction is critical for informed investment.