Team
The model behind CoinSec's first bug finder is now a published paper. BreachT5: Ensembling CodeT5+ Models for Multi-Label Vulnerability Detection in Smart Contracts appears in the proceedings of WETSEB 2026, the workshop on software engineering for blockchain at ICSE, the main academic conference in software engineering. Our advisor, Prof. Annibale Panichella of TU Delft, presented it in Rio de Janeiro, Brazil, in April. The proceedings went live on 4 July.
What the paper does
It asks a plain question: can a model trained on code find the bugs in Ethereum smart contracts that rule-based tools miss, without inventing ones that aren't there? We fine-tuned two sizes of CodeT5+ on more than 67,000 real, open-source Ethereum contracts labelled with twelve vulnerability classes, then combined them so the smaller model's eye for rare bugs and the larger model's precision on common ones work together. That ensemble is BreachT5.
Along the way the paper maps which vulnerabilities are easy to spot and which are hard. Bugs with a clear signature in the code, like an unbounded loop or an unguarded withdrawal, are caught almost every time. Bugs that depend on context, like a call whose safety hinges on who controls the target, are where every tool struggles, and where the work ahead is.
How it did
On the held-out test set BreachT5 reached a Macro-F1 of 0.556 and a Micro-F1 of 0.612, ahead of its own component models, ahead of Slither, the most widely used static analyser, and ahead of GPT-5 used as a zero-shot detector. The code, data pipeline and trained models are open, so anyone can check the numbers.
Why it matters for CoinSec
A classifier can tell you a contract looks vulnerable; it cannot prove it. CoinSec today attacks the program and keeps only the exploits that run. The research collaboration with TU Delft continues, now on the Solana programs CoinSec runs on.
Thanks to our co-authors Gregorio Dalia, Andrea Di Sorbo and Corrado Visaggio, and to the workshop for the slot.