Fine-Tuned CodeBERT for Web Threat Classification using Agentic Reasoning

The internet is a vast, open marketplace where information flows freely, but that openness also invites trouble. Every day, web applications—sites that let you shop, bank, or read the news—face attempts to break in. Two of the most common tricks used by attackers are known as SQL injection and cross-site scripting. In simple terms, SQL injection is like slipping a forged note into a database query, tricking the system into revealing secret information or changing records it shouldn't. Cross-site scripting is more like hiding a malicious script inside a comment or a form field, which then runs on the computer of anyone who views the page, often stealing their login details or taking control of their session. To stop these attacks, websites use digital gatekeepers called web application firewalls. For years, these gatekeepers relied on a list of known bad patterns, like a bouncer checking IDs against a list of known troublemakers. However, clever attackers can easily bypass these lists by slightly changing their malicious code—adding a space, changing a letter's case, or hiding it inside a comment—much like a person changing their clothes to avoid recognition.

To solve this, researchers have turned to artificial intelligence, specifically a type of computer program called a transformer. Think of a transformer as a machine that reads text and learns the underlying structure of language, much like a student who reads thousands of books to understand grammar and context rather than just memorizing specific words. Some of these machines are trained on general text, while others, like the one at the center of this study, are trained on both natural language and computer code. The researchers wanted to know if a machine trained on code could spot these web attacks better than machines trained on general text, and whether it could do so quickly enough to be useful in real-time. They also wanted to see if using a shorter, more efficient version of the machine would sacrifice accuracy for speed, a crucial trade-off for systems that must make decisions in milliseconds.

In a recent study, researchers from Westminster International University in Tashkent set out to test these questions with a fresh, rigorous approach. They gathered a massive collection of web traffic data, mixing together thousands of examples of both harmless requests and malicious attacks. To ensure their results were trustworthy, they cleaned the data meticulously, removing duplicate entries and adding a small set of tricky, harmless sentences that looked like attacks to see if the machines would get confused. They then took four different versions of these AI models and taught them to distinguish between good and bad traffic using the exact same rules and data. One of these models was CodeBERT, a system pre-trained on source code, while the others were general-purpose models like BERT and RoBERTa. The goal was to see if the code-trained model had a natural advantage in spotting the strange mix of database commands and web scripts found in these attacks.

The results were remarkably consistent across the board. All four models performed with extraordinary precision, correctly identifying the nature of the traffic in more than 99 percent of cases. When the researchers looked closely at the few mistakes made, they found that the differences between the models were so small that they could easily be attributed to random chance rather than one model being truly superior. The code-trained model, CodeBERT, made seven errors out of more than 6,000 test samples, a performance that was statistically indistinguishable from the other models, which made between eight and twelve errors. This finding challenges the assumption that a specialized model trained on code is inherently better at this specific task than a general-purpose model, at least when both are given the same amount of data and training time. The study suggests that for the job of simply deciding if a request is malicious or not, the choice of model matters less than the quality of the data and the testing method.

However, the study did uncover important differences in how these models behave under pressure. While the accuracy scores were nearly identical, the types of mistakes they made varied. One model was more likely to let a malicious attack slip through, while another was more likely to mistakenly block a harmless request. This distinction is vital for real-world security, as the cost of a false alarm is different from the cost of a breach. Furthermore, the researchers tested how fast these models could work. They found that by shortening the amount of text the models had to process at once, they could cut the time required to train the system by nearly half and reduce the time it took to make a decision by almost 50 percent, all without losing any significant accuracy. This suggests that for lightweight security systems that need to be fast and efficient, there is a clear path to optimization that does not require sacrificing safety.

The researchers also subjected their models to a stress test using a small set of specially crafted, obfuscated attacks designed to fool the system. In this scenario, the models' performance dropped noticeably, with accuracy falling to around 86 percent. This drop highlights a critical reality: a system that looks perfect on standard test data may still be vulnerable to cleverly disguised attacks. The study explicitly notes that the high scores seen in standard testing do not guarantee robustness against evasion techniques. The researchers were careful to state that their work is a first step, focusing only on the binary decision of whether traffic is bad or good, rather than identifying the specific type of attack. They acknowledge that their dataset, while large, is limited to public sources and that their stress test was small.

Ultimately, this research provides a clear, grounded picture of the current state of AI in web security. It demonstrates that while advanced models can achieve near-perfect accuracy in detecting common web attacks, the differences between specialized and general models are often negligible in practice. The true value lies not in finding a single "best" model, but in understanding the trade-offs between speed, cost, and the specific types of errors a system might make. By showing that efficiency can be gained without sacrificing accuracy, the study offers a practical roadmap for building faster, lighter security tools. It reminds us that in the ongoing battle for web safety, the goal is not just to build a smarter machine, but to build a more reliable and efficient one that can keep pace with the evolving tactics of those who seek to break in.