RevEng.AI launches binary analysis models for code

RevEng.AI has launched Mega Bite, a suite of artificial intelligence models for binary analysis. The release includes two proprietary models, WilBERT and Ventris.

The models are designed to analyse compiled software without access to source code, an increasingly important challenge as organisations rely more on third-party and AI-generated software.

Mega Bite is being offered within RevEng.AI's platform as an extension of its BinNet technology. The models are intended to help security teams inspect software that has already been compiled into stripped binaries, where much of the original development context is no longer visible.

Binary analysis

That gap has long challenged software security. Once code reaches production in binary form, security teams can struggle to determine what is actually running, whether vulnerabilities are present, and whether malicious behaviour has been introduced through external code, dependencies or software acquired from third parties.

Traditional decompilers attempt to reconstruct missing source code context from binaries, but the process is often approximate and can produce pseudocode that is difficult to interpret. General-purpose large language models have also struggled with the task because they are not designed specifically for binary-to-source recovery.

WilBERT and Ventris address different parts of that workflow. WilBERT is designed to identify similar code across binary artefacts, while Ventris is intended to explain executables and recover semantically equivalent source code from them.

Both models were trained on more than 50 trillion tokens of code-to-binary pairs gathered from software built in a range of ways. RevEng.AI said the goal was to improve source code recovery while reducing the hallucinations that can arise when general-purpose models attempt the same task.

Benchmark results

In benchmarking, RevEng.AI said the models achieved 94% accuracy on the HumanEval benchmark for source code recovery, compared with 48% for Anthropic's Fable model and 45% for OpenAI's GPT-5.5.

The launch reflects growing concern in cybersecurity over software supply chain risk. As companies incorporate more open-source components, commercial software and code generated with AI tools, the need to verify software after compilation has become more pressing, especially when source code is unavailable.

The issue is particularly acute in malware analysis, incident response and software assurance, where analysts may need to inspect binaries from suppliers, inherited systems or potentially compromised environments. In those cases, the ability to infer what code is doing without access to the original repository can shape decisions on whether to trust, block or investigate software further.

"Software development is accelerating faster than security teams can manually inspect every application and dependency," said James Patrick-Evans, PhD, Founder and Chief Executive Officer of RevEng.AI. "AI has the potential to fundamentally change how organizations can assess software risk, but only if it delivers results they can trust. Establishing that trust is foundational to RevEng.AI, with WilBERT and Ventris strengthening our platform's ability to provide greater confidence in the software organizations build, buy, and deploy."

Scaling security

RevEng.AI said the models run at inference speeds of about 10,000 tokens per second. That performance allows the tools to be used across existing security workflows at scale rather than only in isolated research settings.

Industry analysts have increasingly focused on how AI may change code review and reverse engineering, but accuracy remains a central concern. For security teams, unreliable output can create a false sense of confidence or force analysts back into manual review, limiting any operational gain from automation.

"As AI generates more of the software organizations rely on, knowing what is inside that software before trusting it becomes critical," said Chris Steffen, Vice President of Research at Enterprise Management Associates. "That level of assurance becomes even harder to achieve with third-party code and complex dependency chains. By creating models designed specifically for binary analysis, RevEng.AI gives security teams greater visibility into what actually runs in production and a way to verify software without needing access to the original source code. This is something the industry has needed for a long time."

Founded in 2023, RevEng.AI focuses on analysing compiled software, including closed-source and third-party artefacts, for enterprise and defence customers. Its broader aim is to help organisations understand the software they ship, buy and run before deployment decisions are made.