Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Check Point Research header illustration for the report Infostealer riding the AI Train, showing a masked outlaw driving a steam locomotive whose tender is labeled AI TRAIN across a desert.
Illustration: Check Point Research, header of "Infostealer riding the AI Train," Oct. 23, 2026 (fair use).

Threat Intel

Check Point Says a New Infostealer Called Aokigahara Has a Dedicated Module for Claude, Codex, Cursor and Other AI Coding Tools, Collecting Their Session Files Alongside SSH Keys and Cloud Credentials

The stealer, which may be a newer version of malware once called Kamidere or Stygian, ships each category of stolen data as its own encrypted upload and identifies itself as Microsoft-CryptoAPI/10.0. Why AI assistant tokens now belong in the credential-exposure playbook, and the YARA rule and network markers Check Point published.

Priya Shah, Threat IntelSingapore5 min read

SINGAPORE - A newly observed password stealer is going after the files that keep developers signed in to their AI coding assistants, according to Check Point Research, adding Claude, Codex, Cursor and other AI developer tools to the browsers, wallets and VPN clients that commodity stealers usually raid.

The malware, which Check Point tracks as Aokigahara, "has been spotted in the wild," the researchers wrote in a report published Friday. "An earlier version may have been known as Kamidere or Stygian."

A module for AI sessions

Most of Aokigahara's handlers are familiar: screenshots, clipboard contents, Chromium and Firefox credentials and cookies, desktop and browser crypto wallets, Discord and Telegram data, Steam and other gaming accounts, Outlook profiles, VPN and FTP configurations, and password managers. What sets it apart is a handler named aisessions.

That handler collects authentication files, session data and configuration for Claude Code and Claude Desktop, Codex, ChatGPT Desktop, Cursor, Windsurf, Kiro, Codeium, Visual Studio Code, GitHub Copilot, Google Gemini CLI and JetBrains AI Assistant, along with AWS credentials, single sign-on caches and Amazon Q data. It also queries Windows Credential Manager. The targeted paths include .claude\credentials.json, .codex\auth.json and the .aws\sso\cache folder.

Check Point's table of paths targeted by the aisessions handler, covering Claude Code and Claude Desktop, Codex, ChatGPT Desktop, Cursor, AWS CLI and SSO, Amazon Q, Kiro, Windsurf, Codeium, Visual Studio Code, GitHub Copilot, Gemini CLI and JetBrains AI Assistant.
Screenshot: Check Point Research, paths targeted by the aisessions handler, Oct. 23, 2026 (fair use).

A separate ide handler pulls Git credential files, files under .ssh that may contain private keys, and AWS, Azure and Google Cloud credential and configuration files from editors such as VS Code, Cursor, Windsurf and JetBrains products. Check Point says not every targeted file necessarily contains secrets, and that "the analysis does not confirm how its operators monetize the stolen data." Valid tokens, it notes, could allow unauthorized use of paid AI services, resale of accounts or exposure of conversations and source code.

Built to avoid analysts

The stealer stops if the keyboard layout or Windows UI language matches any of 11 locales, including Russian, Ukrainian, Belarusian, Kazakh and Uzbek, when that check is enabled. A second flag turns on debugger, virtual machine and sandbox checks, including a list of suspicious user and computer names such as "sandbox", "malware" and "WDAGUtilityAccount". Its configuration sits encrypted in a section named .stgcfg and holds the command-and-control URL, the server's public key, a build ID, a gate token, a module bitmask, optional TLS certificate pinning and up to three fallback URLs.

A distinctive upload pattern

Aokigahara talks to its server through three endpoints under /gate. It opens a session at /gate/handshake by sending a 32-byte X25519 public key, sends each module's loot as a separate AES-256-GCM encrypted ZIP to /gate/chunk with headers that name the module, such as aisessions, and closes with an eight-byte COMPLETE marker to /gate/complete. The encryption key is derived with SHA-256 and the fixed string aokigahara. The latest samples Check Point analyzed identify themselves with the User-Agent Microsoft-CryptoAPI/10.0 and send custom headers including X-Build-ID, X-Gate-Token, X-Session, X-HWID and X-Chunk.

Check Point's table of the three /gate endpoints and an example /gate/handshake request with the Microsoft-CryptoAPI/10.0 user agent and X-Build-ID, X-Gate-Token, X-Session, X-HWID and X-KEX headers.
Screenshot: Check Point Research, Aokigahara network protocol, Oct. 23, 2026 (fair use).

The report lists 150 SHA-1 hashes, a clipper component that is dropped beneath C:\Microsoft\Vault, and command-and-control addresses including 2[.]26[.]252[.]48:8080, dfwioeiofwr[.]info, dobeqnesmot[.]cc, kkeyyou[.]cc and neihite[.]cc.

The YARA rule Check Point published for Aokigahara, matching strings such as X-Chunk-Index, X-Build-ID, AKGRES, AISessions/, aokigahara and kamidere.
Screenshot: Check Point Research, YARA rule for Aokigahara, Oct. 23, 2026 (fair use).

"The key finding is that AI application credentials and session data are part of this malware's collection scope and should be treated as sensitive assets alongside browser sessions, cloud credentials, and developer keys," the researchers concluded.

What defenders can do

  • Add AI coding assistant tokens and session files to infostealer response runbooks: on a confirmed infection, sign the user out of Claude, Codex, ChatGPT, Cursor, Copilot, Gemini and similar tools, and revoke their API keys alongside browser sessions, SSH keys and cloud credentials.
  • Alert on non-developer processes reading .claude\credentials.json, .codex\auth.json, .aws\credentials, .aws\sso\cache or editor globalStorage databases.
  • On the network, look for POST requests to /gate/handshake, /gate/chunk and /gate/complete with Content-Type application/octet-stream and a Microsoft-CryptoAPI/10.0 user agent from processes other than Windows components.
  • Run Check Point's YARA rule across endpoint and sandbox collections, and block the published hashes and command-and-control addresses.
  • Prefer short-lived tokens and single sign-on for AI developer tools so a stolen session file expires quickly.

Sources:


Priya Shah covers threat intelligence, intrusion analysis, and adversary tradecraft for SOCtember from Singapore.

Related stories

Threat Intel desk