A team of researchers has shown they can watch a computer chip think without ever touching it. Using terahertz radiation, a band of the electromagnetic spectrum that slips through materials opaque to visible light, scientists at the University of Adelaide bounced invisible waves off a running processor and read its internal activity from the reflected signal. The technique worked without cracking open the chip’s casing or attaching any probes, and the results, published in IEEE Transactions on Information Forensics and Security in early 2025, have prompted the researchers themselves to flag a troubling possibility: the same method that helps engineers inspect chips could let an attacker silently extract sensitive data.
What the experiment actually showed
The study was led by Prof. Withawat Withayachumnankul and Dr. Chitchanok Chuengsatiansup at the University of Adelaide, working alongside Virginia Diodes Inc., a commercial terahertz hardware supplier, the Hasso Plattner Institute, and the University of Potsdam in Germany. According to the university’s newsroom, the system directs a terahertz beam at a packaged, operating chip and captures the reflected signal. Changes in the chip’s electrical state produce measurably different reflection patterns.
The published results confirm that the system can distinguish between at least two broad computational states: an idle processor and one actively running encryption routines. The researchers have not disclosed the specific chip architecture tested, nor have they published the spatial or temporal resolution their setup achieved. Those omissions make it difficult to judge how granular the technique’s view of internal chip activity really is, and whether it could differentiate between closely related operations rather than just coarse workload categories.
Terahertz radiation occupies the slice of spectrum between microwaves and infrared. It penetrates many plastics, ceramics, and semiconductor packaging materials that block optical inspection. That property has made it valuable for non-destructive testing in manufacturing for years. Earlier work established that terahertz emission microscopy could pinpoint faults inside integrated circuits, though those older methods typically required femtosecond laser excitation and were confined to failure analysis on static chips, as documented in prior microscopy research on large-scale ICs.
The Adelaide study breaks from that earlier generation in a critical way: instead of actively exciting a chip with a laser, it passively reads reflected terahertz signals from a processor running normal workloads. The reflected pattern shifts depending on what the chip is doing, giving an external observer a window into its computational state. Prof. Withayachumnankul has described the work as offering “a powerful non-destructive testing method for chip manufacturers” while also opening “a side-channel attack vector that the security community has not previously had to defend against at this frequency range.” Both he and Dr. Chuengsatiansup have spoken publicly about the dual-use nature of their findings.
The verified takeaway is narrow but important. Information about a chip’s internal logic activity leaks into the terahertz band and can be picked up externally, through the packaging, without physical contact. That is enough to establish a genuinely new class of side-channel signal, even though no one has yet demonstrated a complete attack that recovers, say, a cryptographic key.
Why the real-world threat is still an open question
Proving that terahertz reflections carry information about chip activity is not the same as proving an attacker can weaponize that information. Several practical barriers stand between the lab demonstration and a plausible espionage scenario.
Speed and resolution. Modern processors operate at gigahertz clock speeds. Capturing enough terahertz data, fast enough, to reconstruct high-entropy secrets like encryption keys is a formidable signal-processing challenge. A 2020 study on single-pixel terahertz detection published in Nature Communications demonstrated video-rate acquisition, but only by using computational reconstruction techniques that sacrifice spatial resolution for speed. That work is now roughly six years old, and while subsequent research has continued to push terahertz imaging throughput, no publicly available follow-up has demonstrated the combination of speed, resolution, and sensitivity that would be needed to extract meaningful data from a modern processor in real time. Whether the Adelaide team’s reflected-probing approach can scale to the bandwidth required for meaningful data extraction has not been independently confirmed.
Equipment cost and size. The experiment used laboratory-grade terahertz hardware, including components from Virginia Diodes Inc. That gear is expensive, bulky, and not something an attacker can slip into a backpack. Terahertz sensing circuits fabricated in standard CMOS processes are advancing, and integration trends could eventually shrink the cost and footprint. But “eventually” is doing heavy lifting in that sentence. Today, mounting this kind of attack would demand physical proximity to the target and access to specialized instruments largely confined to research labs and defense agencies.
Chip packaging. Production processors sit beneath heat spreaders, multi-layer substrates, and metal lids that attenuate terahertz signals to varying degrees. The Adelaide researchers have not published detailed results on how different packaging configurations affect signal fidelity. A bare die on a lab bench and a server CPU bolted into a rack-mount chassis are very different targets, and variations in materials and thicknesses could dramatically reduce what an external observer can see.
Environmental noise. A controlled lab lets you isolate a single chip, align the beam precisely, and average out noise. A real data center is full of competing electromagnetic sources, mechanical vibrations, and temperature fluctuations. Each of those factors could blur the subtle reflection differences that encode a chip’s internal states, making practical exploitation far harder than the proof of concept suggests.
Weighing the sources
The strongest evidence here is the peer-reviewed study itself, published in IEEE Transactions on Information Forensics and Security with named institutional affiliations, and the University of Adelaide’s press release identifying the collaborators and quoting the lead researchers. These are primary sources, and they confirm the technique works under controlled conditions. The specific journal title and DOI have not been independently verified by this publication; readers seeking the paper should search the IEEE Xplore database using the authors’ names.
A comprehensive 2023 review of terahertz imaging throughput challenges, published in Light: Science & Applications, provides independent context. That survey catalogs the hardware and computational bottlenecks limiting how fast terahertz systems can capture phase, intensity, and time-of-flight data. It does not address the Adelaide technique specifically, but it frames the engineering constraints any real-world deployment would face.
Readers should be careful to separate what was demonstrated from the security narrative building around it. The researchers proved that terahertz reflection carries information about a chip’s internal state. They did not demonstrate a full attack chain that pulls an AES key from a running server. The gap between “we can see the chip doing something” and “we can read your secrets” is real, and bridging it would require additional signal processing, noise cancellation, and pattern-matching work that has not been shown publicly.
Why security researchers are paying attention anyway
The history of side-channel attacks offers a reason not to shrug this off. Power analysis started as an academic curiosity in the late 1990s before becoming a practical threat that forced hardware and software countermeasures across the semiconductor industry. Electromagnetic emanation attacks followed a similar arc. Acoustic side channels, once dismissed as exotic, have been used to recover RSA keys from the sound of a laptop’s CPU. In each case, the progression from “interesting but impractical” to “we need to defend against this” took years, not decades.
The Adelaide study adds terahertz reflections to that catalog of potential leakage paths. Even if the current implementation is slow, expensive, and limited to short range, it demonstrates a physical principle that cannot be patched away in software. Once a new form of leakage is established, the research community, and adversaries, tend to search for cheaper, more scalable ways to exploit it.
For chip manufacturers, the near-term opportunity is clear: terahertz reflected imaging could improve non-invasive testing and failure analysis, catching defects that other inspection methods miss. For security teams, the study is a signal to begin modeling worst-case scenarios and evaluating mitigations such as modified packaging materials, additional terahertz-band shielding, or deliberate noise injection into the chip’s electromagnetic profile. Standards bodies and policymakers may want to track the technology’s trajectory, though as of May 2026, no regulatory action specific to terahertz side channels appears imminent.
A new way of seeing is also a new way of attacking
The Adelaide work lands on a fault line that runs through all of hardware security: every advance in non-destructive inspection is, by definition, an advance in non-destructive surveillance. The evidence published so far supports the scientific claim that terahertz reflections reveal internal chip activity through sealed packaging. It does not support more dramatic claims about effortless, remote extraction of secrets from everyday devices.
As further experiments test different processor architectures, packaging types, and imaging configurations, the balance between diagnostic promise and security risk will sharpen. The prudent response is neither complacency nor alarm, but close attention to how quickly the gap between a lab demonstration and a fieldable capability closes. If the history of side channels is any guide, that gap tends to narrow faster than defenders expect.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
