A newly documented subscription-based infostealer called Kynx has been observed shipping with a module built for stealing data from AI development tools, according to analysis published on Friday by the SOCRadar Threat Research Unit.
The module, labelled DevGrabber in the malware’s own strings, pulls application data and stored state from eight tools: Claude Code, Cursor, GitHub Copilot, Windsurf, Codeium, OpenAI’s Codex, the ChatGPT desktop app and VS Code. For several, the target is the local state database that holds session and workspace data.
It sits alongside a familiar shopping list of 65 browser extension wallets, 19 desktop wallet paths, 16 gaming platforms and nine VPN clients, plus a sweep of user folders for files whose names suggest wallet recovery phrases.
SOCRadar traced Kynx to a January 2026 advertisement on a Turkish game-modding forum, sold in tiers with the top features gated behind an invite code. The seller claimed to have used Google’s Gemini to help write it.
On execution it displays a fake system update screen. SOCRadar said the staging suggested delivery through cracked software or a ClickFix lure but did not confirm how Kynx reaches victims.
Credential Theft and Chunked Exfiltration
For browser credentials, Kynx runs a cascade of four techniques, moving to the next each time one fails. The first three download public post-exploitation tools that defeat Chromium’s App-Bound Encryption, the protection added to stop this kind of theft.
If all three fail, it copies the raw encrypted databases off the machine and leaves the operator to break them offline.
Session cookies for Instagram, TikTok, X and Spotify are saved with usage instructions attached, so an operator can replay a session rather than crack a password.
Exfiltration runs over ordinary web requests, split into 13 labelled categories covering passwords, cookies, card data, wallets, gaming accounts and the developer haul. Each carries a single-use token valid for five minutes, and the panel permanently blocks any address sending an invalid one.
SOCRadar found the anti-analysis checks failed to fire as designed when the sample was run dynamically, and advised teams to hunt infostealer logs for accounts already exposed.
