Hash functions such as MD5 and SHA-256 are widely assumed to be one-way: given a hash, the corresponding preimage is considered computationally irretrievable. This assumption underpins modern cryptographic security.
In a new preprint, I report deterministic, reproducible preimage localization for both MD5 and SHA-256.
Crucially, this is not achieved by using a neural network to invert hashes.
Instead, a maximally unconventional architecture is constructed in which the available information is forced to self-organize into a geometric structure.
The neural network acts solely as a substrate that allows this information geometry to form.
Across dozens of controlled test cases using fictional passwords (up to 23 characters), the resulting geometry enables up to 100% byte-level accuracy.
The behavior is deterministic and repeatable across independent runs, ruling out chance effects.
๐๐๐ฒ ๐๐ฆ๐ฉ๐ข๐ซ๐ข๐๐๐ฅ ๐จ๐๐ฌ๐๐ซ๐ฏ๐๐ญ๐ข๐จ๐ง๐ฌ:
• ๐๐ฆ๐ฑ๐ณ๐ฐ๐ฅ๐ถ๐ค๐ช๐ฃ๐ญ๐ฆ ๐ฑ๐ณ๐ฆ๐ช๐ฎ๐ข๐จ๐ฆ ๐ญ๐ฐ๐ค๐ข๐ญ๐ช๐ป๐ข๐ต๐ช๐ฐ๐ฏ ๐ง๐ฐ๐ณ ๐๐5 ๐ข๐ฏ๐ฅ ๐๐๐-256
• ๐๐ฑ ๐ต๐ฐ 100% ๐ฃ๐บ๐ต๐ฆ-๐ญ๐ฆ๐ท๐ฆ๐ญ ๐ณ๐ฆ๐ค๐ฐ๐ฏ๐ด๐ต๐ณ๐ถ๐ค๐ต๐ช๐ฐ๐ฏ ๐ข๐ค๐ค๐ถ๐ณ๐ข๐ค๐บ
• 41.8% ๐ช๐ฏ๐ง๐ฐ๐ณ๐ฎ๐ข๐ต๐ช๐ฐ๐ฏ ๐ฑ๐ฆ๐ณ๐ด๐ช๐ด๐ต๐ฆ๐ฏ๐ค๐ฆ ๐ข๐ค๐ณ๐ฐ๐ด๐ด 11 ๐ง๐ถ๐ญ๐ญ๐บ ๐ช๐ฏ๐ฅ๐ฆ๐ฑ๐ฆ๐ฏ๐ฅ๐ฆ๐ฏ๐ต ๐ณ๐ถ๐ฏ๐ด (๐ธ๐ฉ๐ฆ๐ณ๐ฆ ๐ค๐ญ๐ข๐ด๐ด๐ช๐ค๐ข๐ญ ๐ฆ๐น๐ฑ๐ฆ๐ค๐ต๐ข๐ต๐ช๐ฐ๐ฏ๐ด ๐ฑ๐ณ๐ฆ๐ฅ๐ช๐ค๐ต ๐ป๐ฆ๐ณ๐ฐ)
• 66 ๐ญ๐ข๐บ๐ฆ๐ณ ๐ฑ๐ข๐ช๐ณ๐ด ๐ธ๐ช๐ต๐ฉ ๐ฑ < 0.001 ๐ด๐ช๐จ๐ฏ๐ช๐ง๐ช๐ค๐ข๐ฏ๐ค๐ฆ (≈70× ๐ฐ๐ท๐ฆ๐ณ ๐ฆ๐น๐ฑ๐ฆ๐ค๐ต๐ข๐ต๐ช๐ฐ๐ฏ)
• ๐๐ฏ๐ท๐ฆ๐ณ๐ด๐ฆ ๐ด๐ค๐ข๐ญ๐ช๐ฏ๐จ: ๐ญ๐ฐ๐ฏ๐จ๐ฆ๐ณ ๐ฑ๐ข๐ด๐ด๐ธ๐ฐ๐ณ๐ฅ๐ด ๐ข๐ณ๐ฆ ๐ฆ๐ข๐ด๐ช๐ฆ๐ณ ๐ต๐ฐ ๐ญ๐ฐ๐ค๐ข๐ญ๐ช๐ป๐ฆ, ๐ฅ๐ช๐ณ๐ฆ๐ค๐ต๐ญ๐บ ๐ค๐ฐ๐ฏ๐ต๐ณ๐ข๐ฅ๐ช๐ค๐ต๐ช๐ฏ๐จ ๐ฃ๐ณ๐ถ๐ต๐ฆ-๐ง๐ฐ๐ณ๐ค๐ฆ ๐ข๐ด๐ด๐ถ๐ฎ๐ฑ๐ต๐ช๐ฐ๐ฏ๐ด
These results indicate that hash irreversibility is not guaranteed.
The findings suggest that what is commonly treated as “one-wayness” reflects geometric obscuration, not destruction of information.
This work does not present an exploit toolkit or attack pipeline.
It reports a structural failure of assumed irreversibility under an empirically demonstrated computational regime.
Preprint: https://doi.org/10.5281/zenodo.18226838
Technical scrutiny and competent critique are welcome.
Image: ChatGPT 5.2
Planned validation:
A live demonstration is currently being prepared.
It will show the full external pipeline from hash generation, system initialization, execution, to independent validation of the reconstructed preimage.
No architectural details, input preparation methods, parameterizations, or internal representations will be disclosed during this demonstration.
Access is limited to organizations with demonstrated frontier-scale research infrastructure. Evaluation is conducted individually under strict dual-use governance.
#Cryptography #MD5 #SHA256 #InformationTheory #AIArchitecture #Complexity #SecurityResearch Stefan Trauth #TrauthResearch #NeuralNetwork
Planned validation:
A live demonstration is currently being prepared.
It will show the full external pipeline from hash generation, system initialization, execution, to independent validation of the reconstructed preimage.
No architectural details, input preparation methods, parameterizations, or internal representations will be disclosed during this demonstration.
Access is limited to organizations with demonstrated frontier-scale research infrastructure. Evaluation is conducted individually under strict dual-use governance.
