무슨 일이 일어났나요?
IT Brief Australia reports that RevEng.AI launched Mega Bite, a binary-analysis extension for its BinNet platform. The release includes WilBERT, which identifies similar code across binary artefacts, and Ventris, which explains executables and attempts to recover semantically equivalent source code. RevEng.AI says the models were trained on more than 50 trillion code-to-binary tokens and run at about 10,000 tokens per second. The report does not independently confirm those claims or document public access, pricing, or licensing.
IT Brief Australia reports that RevEng.AI launched Mega Bite as an extension of its BinNet technology. The package contains two proprietary artificial-intelligence models: WilBERT, designed to identify similar code across binary artefacts, and Ventris, designed to explain executables and recover semantically equivalent source code. The models are intended for compiled software, including stripped binaries where original development context is no longer visible.
According to IT Brief Australia, RevEng.AI says the models were trained on more than 50 trillion tokens from code-to-binary pairs produced through different software-building methods. The company also says Mega Bite operates at speeds of about 10,000 tokens per second. The report describes potential uses in malware analysis, incident response, software assurance, and review of supplier or inherited systems.
IT Brief Australia reports a claimed 94% HumanEval accuracy for the models’ source-code recovery, compared with figures of 48% for Anthropic’s Fable and 45% for OpenAI’s GPT-5.5. The report quotes RevEng.AI chief executive James Patrick-Evans and Enterprise Management Associates analyst Chris Steffen. No public test results, evaluation methodology, customer examples, pricing, or access terms are provided in the source.
왜 중요한가요?
Binary analysis is important for reviewing third-party, closed-source, inherited, or potentially compromised software when original source code is unavailable. If the reported capabilities perform reliably outside curated benchmarks, they could help security teams investigate supply-chain risks, malware, and production software more quickly. However, the article presents RevEng.AI’s performance figures and comparisons as company-reported results, with no independent testing or methodology detailed.
The launch addresses a specific security problem: analysing software after compilation when the repository, symbols, dependencies, or original developer intent may be unavailable. General-purpose language models and conventional decompilers can produce incomplete or difficult-to-interpret results, according to the report. A specialised system could reduce the time analysts spend triaging unfamiliar binaries, provided its output is treated as investigative assistance rather than proof of safety.
The practical value depends on reliability in adversarial conditions. Malware authors can use packing, obfuscation, unusual compilation settings, and runtime behaviour that static binary analysis may not capture. The source gives no independent evidence about these cases, nor does it establish that recovered code is equivalent in security-relevant behaviour. The reported comparison therefore indicates a company claim, not a confirmed operational advantage.
대화형 메커니즘: 실제로 작동하는 방식
이 개발의 이면에 있는 기본 기술을 대화식으로 살펴보세요.
Which component of an AI application is the machine-learning model itself?
다음에 무엇을 볼 것인가
The key questions are whether Mega Bite is available to customers, how it integrates with existing security workflows, and whether independent evaluations reproduce the reported results. Security teams should also watch for evidence about false positives, missed malicious behaviour, handling of stripped or obfuscated binaries, data handling, and the limits of recovered source code.
The source does not say whether Mega Bite is generally available, limited to enterprise or defence customers, offered through an API, or restricted to selected users. It also does not provide pricing, deployment options, data-retention terms, or information about whether customer binaries leave an organisation’s environment.
Further reporting should examine independent tests on stripped, optimised, obfuscated, and malicious binaries; rates of incorrect similarity matches; the quality of recovered explanations; and how analysts validate the models’ conclusions. Evidence from deployments in incident response or software-assurance workflows would be more informative than results alone.
Security teams considering the product should seek clear documentation on supported architectures, file types, privacy controls, auditability, human review requirements, and how the system signals uncertainty. None of those access or operational details is established by the source.