You may have heard this before: when reading words in a familiar language, it hardly matters in which order the letters appear as long as the first and last letters are in the correct place. The human brain automatically corrects the chaos while reading.
Try it yourself:
Weher cuold one go on a smuemr vaitcoan, if one wtnas to raelx, lkies ncie wtaeher, and at the smae tmie dsiocver smoehitng new?
The original, of course, is:
Where could one go on a summer vacation if one wants to relax, likes nice weather, and at the same time discover something new?
It is hardly surprising that such a sentence poses no hurdle for modern AI. However, this capability of language models was already observed back in November 2023 and it was quite surprising at the time.
Researchers then went a step further and scrambled not just the middle letters, but the entire content of the words:
ehWer doluc eno og no a suemrm tncaoavi if eno tnsaw ot lrxea, isekl icen hetraew, nda ta eht ames emit oicdvres thgmeniso nwe?
Even here, language models had almost no difficulty. Often, they didn’t even explicitly notice the errors, but instead replied directly to the content of the sentence.
Without insight into the exact training data, it is hard to say whether the models were specifically trained with "noisy" text or if this stability is an emergent phenomenon of massive datasets. The latter is more likely: to understand scrambled text, the model doesn't need to calculate the exact original string; it merely needs to extract enough signals from word form and context to reconstruct the most likely meaning. This is exactly where Large Language Models (LLMs) are unbeatable.
One thing is certain: AI couldn't care less about typos in your chat messages.
A lot has happened since those first observations. We have entered the era of "Reasoning" and autonomous agents. How far has technology come today?
The following stress test provides an impression: delete all spaces from the already scrambled text:
ehWerdolucenoognoasuemrmtncaoaviifenotnsawotlrxea,iseklicenhetraew,ndataehtamesemit oicdvresthgmenisonwe?
If you give this block to a modern AI agent and ask for an analysis, without mentioning the language or how the chaos came about, something amazing happens. The analysis takes significantly longer, but the result is correct.
For a human, deciphering this sentence without prior knowledge of the encryption or language would be nearly impossible. The AI manages it in a few minutes. This proves that in analytical tasks within the technical limits of these models (such as context size), they have long since become superhuman.
This talent has not gone unnoticed in the developer community. Beyond pure coding, the AI’s strength lies primarily in finding subtle bugs in source code. For example, over a hundred bugs, 22 of them security-related, were recently discovered in the Firefox browser solely through the analysis of modern models. Also Anthropic has not yet released its strongest model, Mythos, because it believes the model is too powerful at finding bugs. The company wants to give widely used applications a head start, and it has already identified thousands of high-severity vulnerabilities, including some in every major operating system and web browser.
But the AI truly flexes its muscles when analyzing Legacy Code. This old code was often written at a time when "software architecture" was a foreign concept and storage space was extremely limited. The consequences:
Often, the only real documentation existed in the head of the developer, who has long since retired. The problem is gargantuan: estimates suggest that more than 200 billion lines of COBOL still control our global lives in the background. From banking systems to healthcare and retail.
It is precisely this pattern recognition within chaos, contextual reconstruction, and semantic understanding that can work wonders here. The demand is now so high that the first startups and products are emerging. And in our own software modernization projects, AI is now allowed to utilize its full potential.
However, legacy code is just the precursor to an even more difficult problem: what happens when the code itself is no longer available?
Normally, humans write code in a high-level language. This code is readable and contains function names and comments. Before a computer can execute this program, it is compiled. In this process, the readable text is translated into binary machine code. During this transition, almost everything helpful to humans is lost: variable names disappear, structures are flattened, and logic is reduced to a minimum.
The work of reverse engineering is like a puzzle with a million pieces that also looks like an endless phone book. These are tasks that require immense knowledge and endurance. Until now, this was painstaking manual labor for specialists using tools like decompilers to read cryptic assembly instructions. But this is exactly where the new AI agents are stepping in.
Because this is a very niche topic, there are few third-party reports so far. Therefore, I can only draw on my own experience as a reverse-engineering hobbyist. In my spare time, I occasionally try to analyze old computer games to bring them back to life. The source code has often been lost for decades. This work is tedious, but satisfying once you decode a piece of this cryptic text from the distant past.
With coding agents, we can now speak of a "Renaissance." In analytical questions, AI is superior to me by a factor of roughly 100:1. It was able to unlock everything under 3,000 lines of code within a few hours. Personally, I would have needed several weeks for the same task:
On the left, you see the cryptic starting point. The agent is given 3,000 lines of this. On the right, in the comments, is the agent's analysis of the code after several hours of work.
Of course, AI often lacks context, and many of its assessments remain general at first. That’s exactly where I come in to help it along.
I also tested the agent by providing cryptic files from the last four decades from my hard drive with the simple question: "What is this file? Can you decode it?" Even here, the agent often identifies the encoding effortlessly and writes suitable extractors.
The capabilities of these early agents suggest to me that reverse engineering will no longer be a fundamentally unsolved problem in the software industry in the long run. While the task is highly complex, it has one decisive advantage: the result is relatively easy to verify and thus trainable. The only thing that matters is whether the reconstructed code behaves like the original. Because finished code can be compiled and given to the AI to reconstruct, there is practically unlimited training material available.
It is also remarkable that these agents have not been explicitly trained for reverse engineering. The vast amount of existing, often mediocre code seems to provide enough structure for transfer learning to take hold.
Of all the new capabilities of AI agents, this is the one that has impressed me the most: recognizing hidden structures in massive, cryptic datasets was long considered a deeply human strength. Yet, this is exactly where AI agents are now pulling ahead of us. Quietly, and almost unnoticed.