# AI-assisted reverse engineering (OpenAI Codex + Binary Ninja MCP) defeats encryption in NetacFileLock, SecZure YutianDun, KakaSoft USB Copy Protection and Nexcopy Copy Secure USB drive software

> Researcher Xusheng Li used an AI coding agent (referred to as Codex) driving Binary Ninja through its Model Context Protocol interface to reverse four USB drive encryption/copy-protection products in roughly 7-36 minutes each. The analysis recovered embedded fixed keys, built-in backdoor account passwords and weak key derivation that allow decryption without the user password. No CVEs, vendor responses or in-the-wild exploitation are reported.

- **Published:** 2026-10-26T07:21:06Z
- **Last reviewed:** 2026-10-26T07:21:06Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-4434
- **ID:** TL-2026-4434
- **Severity:** MEDIUM
- **Category:** VULNERABILITY
- **Status:** TRACKING
- **Detections:** 9 · **IOCs:** 14 (full data via the Threadlinqs MCP server — Purple tier)

## Description

On 2026-10-22 Xusheng Li published 'Breaking more encrypted USB drives with AI' on the Binary Ninja blog, extending earlier manual research (a time-travel-debugging talk and the Paged Out! issue 5 article 'Examining USB Copy Protection', November 2024). GBHackers covered it on 2026-10-26. An AI agent connected to Binary Ninja via MCP unpacked installers, recognised cryptographic code patterns, reverse engineered a custom hash function and recovered key material statically, including analysis of Windows binaries from macOS. The AI reproduced the earlier NetacFileLock finding in about 7 minutes and found new design flaws in three previously unanalysed products.

NetacFileLock: a fixed 32-byte AES-256 key is embedded in the executable; the user password is checked separately (MD5-based verification) and does not determine the encryption key. Files are encrypted ECB-style in 16-byte blocks with no IV or chaining, and in 'fast mode' only the first 1,024 bytes of supported formats (PNG, JPEG, GIF, TIFF, DWG, PSD, PST, DBX, Office, MDB, WPD, PDF, ZIP, RAR, WAV, AVI, RAM, RM, MPEG, MOV, ASF, MIDI) are encrypted. File signatures use unkeyed MD5, an SM4 option uses another fixed key, and passwords are written in plaintext to diagnostic logs.

SecZure YutianDun (AQUP): credentials are hashed with MD5 and keys wrapped with HMAC-MD5, with AES-128 in ECB mode for file encryption. Each AES key uses 16 hex characters, capping entropy at 64 bits. Three accounts can unwrap the content identifier and decrypt files: Buffer/123456, System/SECZURE.COM (the primary post spells it SERZURE.COM in one place; sources disagree) and the administrator with a user-set password. The built-in accounts bypass the user password entirely.

KakaSoft USB Copy Protection 5.6.0.299: the Normal setting only hides/locks files with no encryption. High XORs the first 16,385 bytes with a single-byte key (256 possibilities; known file signatures reveal it). Highest uses 256-bit Blowfish (layout: 8-byte salt, 8-byte IV, ciphertext, 75-byte trailer) but the key derives from a modified MD5 (altered round-three rotation counts 4,11,16,16,4,11,11,4 and permuted hex alphabet 1903642857ACBEDF) of an embedded fixed seed plus RIPEMD-128(salt || internal password), enabling passwordless decryption. Files are hidden in an obfuscated exFAT tree using reserved device names (COM1) and Unicode private-use characters (U+E347).

Nexcopy Copy Secure (provisioned with Drive Wizard 1.0.81): files use three-key Triple-DES in CBC mode, but a fixed 32-byte certificate embedded in the Mac and Windows viewers decrypts a configuration file that holds the 24-byte 3DES key, 8-byte IV and a hardcoded password (123456). The Mac viewer compares the password with a plain string comparison, so the password only gates the viewer UI. Anti-capture checks (process enumeration, screenshot and clipboard monitoring, detection of ScreenFlow, Camtasia, Debut, Grab and QuickTime) are bypassable.

The author deliberately skipped coordinated disclosure, arguing that issued devices cannot be recalled or updated. The sources report no CVE assignments, CVSS scores, vendor responses or exploitation in the wild; severity is an analyst estimate. Defender takeaway: password prompts and copy restrictions in these products do not provide cryptographic confidentiality, and AI-assisted static analysis lowers the cost of finding such flaws in other legacy encryption tools. Note: the hunt skeleton and GBHackers attribute the work to OpenAI Codex, while the primary post refers only to 'Codex' and 'the AI' without naming a model.

## MITRE ATT&CK

- T1078.001 Default Accounts
- T1078.001 Default Accounts
- T1140 Deobfuscate/Decode Files or Information
- T1552.001 Credentials In Files
- T1552 Unsecured Credentials
- T1005 Data from Local System
- T1025 Data from Removable Media
- T1113 Screen Capture

## Sources

- [Breaking more encrypted USB drives with AI (Binary Ninja blog, Xusheng Li)](https://binary.ninja/2026/10/22/breaking-more-encrypted-usb-drives-with-ai.html)
- [OpenAI Codex Breaks USB Drive Encryption in Minutes, Exposing Critical Security Flaws (GBHackers)](https://gbhackers.com/openai-codex-breaks-usb-drive-encryption-in-minutes/)
- [Breaking Encrypted USB Drives with Time-Travel Debugging (slides, xusheng6)](https://github.com/xusheng6/breaking_encrypted_USB_drives_with_TTD/blob/main/Breaking_Encrypted_USB_Drives_v2.pdf)
- [Examining USB Copy Protection, Paged Out! issue 5](https://pagedout.institute/download/PagedOut_005.pdf)
- [Xusheng Li (xusheng6) GitHub profile](https://github.com/xusheng6)
- [Binary Ninja MCP server (fosdickio)](https://github.com/fosdickio/binary_ninja_mcp)
- [Binary Ninja](https://binary.ninja/)

## Full data

Detection queries (Splunk SPL / Microsoft KQL / Sigma) and IOC values require the Threadlinqs MCP server (Purple tier): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-4434