Breaking

RevEng.AI releases Mega Bite models to analyze stripped binaries without source code

September 3, 2026 at 08:51 EDT

Foundation Models
Security

RevEng.AI has launched Mega Bite, a suite of specialized AI models built to inspect compiled software when original source is missing. The September 2, 2026 release adds two proprietary models—WilBERT for function matching and Ventris for neural decompilation—so security teams can analyze third-party, commercial, and AI-generated binaries after compilation has stripped away source context.

September 2, 2026 · RevEng.AI

Source-free binary analysis, at a fraction of frontier cost

Mega Bite pairs WilBERT function matching with Ventris neural decompilation so teams can inspect stripped production binaries without original source.

HumanEval cost gap

Same benchmark, whole-run cost. Column height tracks the multiple; the 1× bar is enlarged so it stays visible.

1×

Ventris

$0.20

97×

Sonnet 5

no thinking · $19.44

579×

Sonnet 5

high thinking · $115.72

Takeaway: Ventris matched frontier-range HumanEval scores at $0.20 for the full run — 579 times cheaper than Claude Sonnet 5 at high thinking.

Function match Hit@1 · AArch64 Linux

38,270 function pairs. Each block is 5 percentage points.

0.790

WilBERT

0.473

Ghidra BSim

0.021

IDA FLIRT

85.4%

Ventris · Linux x86-64 HumanEval

88.4%

Ventris · Windows x86-64 HumanEval

50T

code-to-binary training tokens

10k

tokens per second inference

From source to recovered binary context

Write and assemble

AI-generated and third-party code enter the chain.

→

Compile and strip

Production ships a binary, not the original source.

→

Context disappears

Names, types, and intent are hard to recover by hand.

→

Match and recover

WilBERT finds similar code; Ventris reconstructs source.

What practitioners welcome

Specialized models land in the same HumanEval band as expensive frontier setups, with stronger function retrieval than FLIRT, BSim, or generic embeddings — and at a small fraction of the token bill.

What remains unfinished

HumanEval tasks are short and stdlib-only. Ventris still lacks full ARM coverage on DecBench. Training-data leakage is possible, and VeriDecomp’s 65.8% recompile rate is preliminary, not a proof of equivalence.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.