Are there hidden structures below the pyramids of Giza?

In 2022, Filippo Biondi and Corrado Malanga published a method that processes a single satellite radar image to reveal subsurface structure. They applied it to the Great Pyramid of Giza and reported imaging internal chambers and features deep below the plateau. A 2025 press conference extended the claim to large structures hundreds of metres underground, and the story went viral. On 10 August 2026 the journal retracted that paper.

Does the published method actually show what the viral claim says it shows?

I could not reproduce evidence that the published single-pass method detects chambers or deep structures beneath the Great Pyramid.

I wanted to. The satellite radar images looked extraordinary, Biondi had encouraged people to check his work, and I got curious enough to try. But when I ran the processing I had implemented on the pyramid and a nearby comparison patch, the pyramid never stood out clearly. I tried 240 variations. A preliminary check across five sites left me with much the same problem: very different places, very similar-looking profiles.

If the processing cannot tell the Great Pyramid apart from the ground nearby, it is hard to know what those convincing-looking structures are supposed to mean.

My own work needed a correction too. The ~285 m elevation cell I originally reported is withdrawn. Splitting one radar image does not create the independent elevation views that standard tomography needs.

The original 2022 paper was retracted on 10 August 2026. My tests cover only part of its processing, and that retraction does not turn them into a full reproduction. I went into this looking for buried architecture. These tests did not give me evidence that I had found it.

Read the corrected paper — Zenodo v2

Millions of people saw the claims. I wanted to check the paper.

Last year, an Italian engineer named Filippo Biondi went viral by claiming he had discovered deep underground structures beneath the Giza pyramids. In January 2026, he went on Joe Rogan and explained his findings in detail. According to his account, there were shafts extending hundreds of meters down and structures at over a kilometer of depth. He showed colorful images, referenced his 2022 peer-reviewed paper with Corrado Malanga, and mentioned the Italian Space Agency.

He also pointed to two validations. First, he said the method correctly detected the Grand Gallery, Queen's Chamber, and King's Chamber inside the Great Pyramid. Second, he said it picked up the Gran Sasso physics laboratory, about 1.4 kilometers inside a mountain. If the method gets the known targets right, the unknown deep structures are supposed to look much more credible.

That is an extraordinary claim. If true, it would rewrite a lot more than pyramid archaeology.

On the show he also encouraged independent verification, since the experiments can be run with openly available data sources. I'm not a radar scientist, I didn't know what SAR was until recently, and I have no university affiliation. But I read the paper, I have a computer, and I got curious enough to try reproducing it.

I published the first version of my write-up in April 2026. The paper I was testing was later retracted on 10 August 2026. My own work also needed corrections, including one to the geometry argument I originally put near the top of the article. Those are included here and in the corrected paper, Zenodo v2.

Can this method actually see the known chambers?

The first thing I checked was the low hanging fruit: whether the method can see the chambers we already know are inside the Great Pyramid.

The method, as described in the paper, is supposed to work from a single satellite radar image. You split that image into Doppler sub-apertures, work with the differences between them, and reconstruct what lies inside or below the structure. The pitch is basically: one image in, underground tomography out.

That sounds amazing. It also creates a very simple question:

Does the image contain the different views you would need to separate one height from another?

Those sub-apertures are slices along the satellite's flight path. They do not give you the independent elevation views used in standard SAR tomography. Cutting the image into more pieces does not add views the satellite never recorded.

I originally tried to put a number on that limitation. That calculation was wrong, and the ~285 m elevation cell I reported is withdrawn. I had treated differences along the flight path as though they supplied an elevation baseline. The correction leaves me without a valid single-pass elevation cell to compare with the chambers.

The King's Chamber is about 43 meters above the pyramid's pavement. The Queen's Chamber is about 21 meters above it. Those are heights inside the pyramid, and separating them requires information that this single-image geometry does not supply under standard tomography.

There is also the question of how the signal gets there. X-band and C-band radar do not penetrate intact limestone far enough to image those chambers directly. The paper's proposed route through surface vibrations would need its own explanation and validation. I have not established that connection.

But maybe I'm missing something. Let's test it.

Maybe there was something about the processing that I was not understanding. So I tried it on real data.

I took two patches from the same Giza radar scene. One was centered on the Great Pyramid. The other was about three kilometers southeast, away from the pyramid itself. Then I ran the same processing on both.

I need to correct something about that second patch too. In the first article I called it empty desert. It falls on settled, irrigated ground at the edge of the plateau. It is still a useful comparison from the same scene, but I should not have described it as empty sand.

What I implemented was the Doppler beamforming part of the idea. It does not include the paper's full processing chain, including its estimation of tiny movements. That matters when interpreting what happened next.

I ran 240 experiments, changing how the image was split, how it was filtered, and how far the reconstruction was asked to look.

In none of those runs did the Great Pyramid stand out clearly from the comparison patch. On average, the other patch produced a slightly stronger response.

That was hard to get past. I had started with the pyramid because it contains known chambers. Yet the processing I was running gave me no clear way to pick it out from the ground nearby.

Whatever those profiles were showing, I could not use them as evidence that I had detected the chambers.

Different places. Very similar profiles.

I still wasn't satisfied though. Maybe something about that one comparison patch was confusing the result.

So I widened the check. Using a stack of 15 Sentinel-1 images from 2023, I looked at five sites in the same scene: the Great Pyramid, the southeast comparison patch, the Sphinx, downtown Cairo, and agricultural land.

The pyramid and the southeast patch again produced almost identical profiles. The other sites looked very similar too. Even at the height of the King's Chamber, the pyramid did not stand out above the comparison patch.

This was a preliminary look at a stack of images. I had not completed all the alignment and terrain corrections that a careful multi-pass analysis needs, so it cannot tell us what a fully processed stack would show.

But it left me with the same question as before. If very different places keep producing very similar-looking structure, how do I know the picture is telling me something about what is actually there?

Why a reconstruction can look like a discovery

The best way I found to explore that question was to build a small example where I already knew the answer.

I planted a few known pyramid heights into a toy model, including the base, the Queen's Chamber, and the King's Chamber. Then I ran a tomographic reconstruction and looked at what came back.

Could shallow features produce convincing-looking patterns much farther down?

In that toy model, yes.

With only a handful of viewing baselines, structured-looking patterns appeared well below the features I had put in. When I gave the reconstruction a much denser set of views, those deep patterns faded and the planted features became clearer.

I knew there was nothing deep in the model because I had built it. The reconstruction had produced something that could easily look meaningful if I had not known what was there.

That does not tell me how Biondi's published images were made. The toy uses its own setup, and I cannot point at one of his figures and say I have reproduced the source of every bright line. What it shows is why a persuasive picture needs a test outside the picture itself.

If you already expect shafts, chambers, spirals, or giant underground structures, it becomes very easy to read those patterns as discoveries.

What about Gran Sasso?

The Gran Sasso laboratory is part of what made the story interesting in the first place. A method that could reliably detect a known laboratory deep inside a mountain would deserve a close look.

I did not run that experiment. There are no Gran Sasso results in my paper.

What I would want to see is whether the processing can pick the laboratory out without relying on someone already knowing where to look, and what it produces in comparable places without a laboratory. A bright feature at an expected depth is a starting point for that check.

My Giza experiments gave me a reason to ask for those comparisons. They did not give me an answer about that mountain.

One image, or a stack?

Up to this point, the question I had been trying to follow was the paper's single-pass claim.

One radar image goes in. A depth reconstruction comes out.

But then I noticed something odd.

The paper is framed around single-pass processing, yet Table 2 lists twelve SAR products. The acknowledgments also thank the developer of SARPROZ, a tool used for multi-temporal radar processing. Neither detail tells me exactly how the figures were produced, but together they leave an obvious question:

Which steps used one image, and which used several?

That distinction matters. Images taken on different passes can provide real differences in viewing position. Those can help separate elevations in a way that slicing up one image does not.

On Rogan, Biondi also said they had run more than 200 scans and got uniform results. I would need to know what those scans were doing. Repeating a single-image procedure and combining different passes into a stack are different experiments.

And even a stack with much better elevation resolution still has to explain how the radar carries information about a chamber through the limestone around it. A sharper reconstruction does not, by itself, make the signal reach farther into the rock.

So what did I actually find?

I started this because I wanted to know whether those underground structures were real. I ended up spending much more time on a smaller question: could I reproduce a result that distinguished the pyramid from somewhere else?

With the processing I implemented, I could not.

I also found that my first geometry calculation needed to be withdrawn, that my comparison patch was not the empty desert I had called it, and that I needed to be much more careful about the distance between testing part of a method and reproducing the whole thing.

The retraction of the original paper does not settle those details for me. It does not prove my calculations or explain the published pictures. My work has to stand on what the experiments actually show.

Those experiments did not give me evidence of chambers or deep structures beneath the Great Pyramid. They showed how readily the processing I tested could produce similar-looking profiles at different places, and how a simple reconstruction could create deep-looking patterns from shallow features.

That is enough to make me cautious about the images. It leaves the larger archaeological question open.

Check my work

Before publishing the first version, I emailed Prof. Biondi. He had publicly encouraged independent verification, and that was what I set out to do.

The code is open source, and the corrected write-up is on Zenodo as v2. The tests and their limits are there for anyone who wants to follow them.

If I'm wrong, please let me know. I really wanted to believe there were underground structures and went into this process looking for them. I still want to know what's there. These tests have not shown me.

This is about how the research actually got done, in a different way from what I planned.

I'm not a domain expert in SAR. I wanted to tell an agent the goal — test these radar claims — and have it figure out the rest. Read the paper, design the experiments, write the code, run it on real data, iterate until there was an answer.

The original setup was OpenClaw running an autonomous loop, with Claude Opus as orchestrator and Sonnet doing the coding. It worked for about a third of the project. Then the context problem showed up.

Research is not a sequence of independent stages. Each decision constrains the next, the whole chain is coupled. When the agent's context got compacted, those couplings broke. Stages got rewritten using conventions that contradicted earlier ones, producing outputs that looked numerically valid but were scientifically meaningless. I tried various OpenClaw memory "hacks", but none of them really worked.

The actual research got done in a single long Claude Desktop conversation. I described what I wanted to test, Opus wrote the script, I ran it on the DGX Spark, pasted the terminal output back, and Opus interpreted, proposed the next experiment, and wrote the next script. Human in the loop. The context was never compacted, which mattered later. When ChatGPT caught a formula error in peer review, Opus could trace it through every downstream calculation and every section of the paper it affected.

One piece did work as an autonomous loop: the 240-experiment parameter sweep. A precise question with an unambiguous evaluation metric — run, score, log, move on. Autonomy belongs to bounded problems. It breaks when the problem itself needs to evolve based on what you are finding.

There's another thing that worked unexpectedly well: multi-model peer review. Four rounds with ChatGPT, Gemini, and Grok, each unaware of the others. Gemini caught a unit error in the main table. ChatGPT caught a formula mistake that moved an early resolution figure from 11.3 km to about 285 m, and a bug where a swept variable was logged but never entered the computation. A later review withdrew that 285 m figure, and the ~618 mm baseline under it. The construction was not an elevation aperture. Grok caught framing weaknesses. Each model was good at different things. Running them independently produced better coverage than any single model doing multiple passes.

The process I trust now: plan first, front-load context, use the agent for bounded well-scoped tasks, run multi-model peer review before anything goes public.

The longer-term goal is still the original one — an agent that can do the full research on its own. That would be a game changer. People could chase the questions they actually find interesting and let agents handle the experiments, which they can run faster and more systematically than any human. We are not there yet, but every day we're getting closer.

Read the full write-up →

Primary target paper

• Biondi, F.; Malanga, C. (2022). "Synthetic Aperture Radar Doppler Tomography Reveals Details of Undiscovered High-Resolution Internal Structure of the Great Pyramid of Giza." Remote Sensing 14(20), 5231. https://doi.org/10.3390/rs14205231 — retracted 10 August 2026.
• Retraction notice. Remote Sensing 2026, 18, 2679. https://doi.org/10.3390/rs18162679

Companion / related Biondi work

• Biondi, F. (2022). "Scanning Inside Volcanoes with Synthetic Aperture Radar Echography Tomographic Doppler Imaging." Remote Sensing 14, 3828. https://doi.org/10.3390/rs14153828. The volcano paper discusses imaging on the order of about 3 km, not a validated 6 km detection.
• Biondi, F.; Addabbo, P.; Clemente, C.; Orlando, D. (2019). "Measurements of Surface River Doppler Velocities..." Remote Sensing 11, 766. https://doi.org/10.3390/rs11070766
• Biondi, F.; Malanga, C.; Mei, A. Press conference, Castel San Pietro Terme, March 15-16, 2025. Subsurface "city" claims; not peer-reviewed.

SAR tomography foundations

• Reigber, A.; Moreira, A. (2000). "First Demonstration of Airborne SAR Tomography Using Multibaseline L-Band Data." IEEE TGRS 38(5), 2142-2152. https://doi.org/10.1109/36.868873
• Fornaro, G.; Lombardini, F.; Serafino, F. (2005). "Three-Dimensional Multipass SAR Focusing." IEEE TGRS 43(4), 702-714. https://doi.org/10.1109/TGRS.2005.843567
• Lombardini, F. (2005). "Differential Tomography: A New Framework for SAR Interferometry." IEEE TGRS 43(1), 37-44. https://doi.org/10.1109/TGRS.2004.838371
• Zhu, X.X.; Bamler, R. (2010). "Tomographic SAR Inversion by L1-Norm Regularization." IEEE TGRS 48(10), 3839-3846. https://doi.org/10.1109/TGRS.2010.2048117
• Zhu, X.X.; Bamler, R. (2010). "Very High Resolution Spaceborne SAR Tomography in Urban Environment." IEEE TGRS 48(10), 4296-4308.
• Fornaro, G.; Pauciullo, A. (2009). "LMMSE 3-D SAR Focusing." IEEE TGRS 47(1), 214-223.

SAR penetration physics

• McCauley et al. (1982). Science 218(4576), 1004-1020. https://doi.org/10.1126/science.218.4576.1004
• Schaber, G.G. et al. (1986). IEEE TGRS GE-24(4), 603-623. https://doi.org/10.1109/TGRS.1986.289677
• Schaber, G.G.; Breed, C.S. (1999). Remote Sensing of Environment 67, 318-332.

Ionospheric / split-spectrum

• Gomba, G. et al. "Toward Operational Compensation of Ionospheric Effects in SAR Interferograms: The Split-Spectrum Method." PDF in project knowledge.

Independent Giza subsurface ground truth

• Morishima, K. et al. (2017). "Discovery of a Big Void in Khufu's Pyramid by Observation of Cosmic-Ray Muons." Nature 552, 386-390. https://doi.org/10.1038/nature24647
• Procureur, S. et al. (2023). Nature Communications 14, 1144. https://doi.org/10.1038/s41467-023-36351-0
• Cardarelli, E. et al. (2025). "Confirmation of the Scanpyramids north face corridor in the Great Pyramid of Giza using multi-modal image fusion..." Scientific Reports 15, 9275. https://doi.org/10.1038/s41598-025-91115-8
• Elkarmoty, M. et al. (2023). "Localization and shape determination of a hidden corridor in the Great Pyramid..." NDT & E International 139, 102809. https://doi.org/10.1016/j.ndteint.2023.102809
• Cardarelli, E. et al. (2025). "Modified ERT cavity detection..." Scientific Reports 15, 41187. https://doi.org/10.1038/s41598-025-29081-4

• Corrected paper on Zenodo (23 October 2026). All versions: 10.5281/zenodo.19574700.
• Reproduction code on GitHub