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.

Hunt.io infrastructure map of the server at 145.63.138[.]67 at Podaon SIA in the Netherlands, showing the HavoX C2 on port 8000, the Phantom53 task broker on port 8080, the Kraken bridge to an external panel, file upload and archive ports, a SOCKS5 proxy, 12 Cloudflare Workers fronting domains and 603 victim implants.
Image: Hunt.io (fair use). Infrastructure map of the operation's single VPS, from the Hunt.io report published Nov. 18, 2026.

Threat Intel

Hunt.io Says an Exposed Server Shows a Russian-Speaking Crew Cracking Victims' Password Vaults Offline and Harvesting ID.me, Login.gov and Bank Logins From 603 Infected PCs

The operator's own open directory held decrypted LastPass and 1Password vaults, Chrome logins pulled past app-bound encryption, Microsoft 365 cookies replayed to mint Graph tokens without an MFA prompt, and scripts that enroll ScreenConnect and Level.io agents for persistence.

Priya Shah, Threat IntelSingapore6 min read

SINGAPORE - A single rented server in the Netherlands, left with an open directory, has exposed the working files of a credential-theft operation that infected 603 Windows computers in less than four weeks and walked off with decrypted password-manager vaults, banking logins and the federal identity accounts Americans use to reach the IRS and Social Security, according to research published Wednesday by the threat-hunting company Hunt.io.

Hunt.io said its AttackCapture service found the directory on a VPS at 145.63.138[.]67, hosted by Podaon SIA, with warnings recorded on Oct. 20 and Nov. 2. The exposed folder on port 8124 held 1,692 files totaling about 1.11 GB: loot archives, post-exploitation scripts, command-and-control components and the operator's logs. The company said implants beaconed from Oct. 12 to Nov. 8.

The researchers stressed that no bank or government agency was breached. The theft happened one step earlier, on victims' own PCs, where many people had a password manager and multifactor authentication in place. In the recovered data, at least 57 distinct victims held a U.S. bank or government login and at least 46 held a federal benefits or identity account, with about 20 ID.me and about 19 Login.gov accounts. Hunt.io described both counts as floors.

Three command servers on one box

The server ran three custom components side by side, Hunt.io said. HavoX, a Python implant manager on port 8000, sat behind a Russian-language login page reading "Authorization Required." Phantom53, a Go binary its author calls "adminka," the Russian slang for an admin panel, brokered tasks from a SQLite database on port 8080. A third Go process, which the operator calls krint and Hunt.io calls the Kraken bridge, relayed tasks to an external panel at 146.70.24[.]172 and served them to implants as single-use download links that expire after 30 minutes.

All victim traffic was fronted by 12 Cloudflare Workers domains. Hunt.io counted 1,242,913 beacon transactions across the fleet, and a separate log listed 603 unique bot IDs on 469 distinct hostnames. The implant itself is a Python .pyw script stored under AppData\Roaming and run by pythonw.exe so no window appears, with a build ID of 4444444 and a protocol value of 0xdeadbeef. The report maps initial access to phishing.

Hunt.io diagram of the attack chain, from initial delivery of the Python implant and C2 beaconing through credential harvesting to parallel modules for LastPass, 1Password and KeePass cracking, Microsoft 365 MFA bypass, RMM persistence and AI tool exfiltration, ending in high-value target processing.
Image: Hunt.io (fair use).The operation's attack chain, with each stage documented by a script or log in the open directory.

Under the password manager

The directory shows why a password manager did not save these victims. One script, dec_chrome20.py, hardcodes AES-GCM keys for Chrome's app-bound encryption formats and picks the right key from the prefix on each encrypted blob, letting the operator decrypt stolen credential and cookie databases offline. Across 16 victims, Hunt.io counted at least 3,345 decrypted Chrome and Edge login records with plaintext passwords.

A seven-script pipeline worked through a LastPass extension store taken from one victim's Chrome profile until it produced 2,880 logins, plus 158 further items that included 17 passkeys and the victim's LastPass sharing RSA private key. Separate scripts confirmed 1Password master passwords for two victims through an offline key derivation crack, and another read a live KeePass database from a developer's workstation without closing the application.

Past the MFA prompt, and into AI tools

For Microsoft 365, the operator did not need passwords. Hunt.io said stolen ESTSAUTHPERSISTENT cookies from two infected machines were replayed through a PKCE authorization request with prompt=none, which returned an authorization code without a login or MFA prompt. A script that went through seven versions turned those codes into live Microsoft Graph bearer tokens, and the files record two confirmed compromises in separate tenants.

Hunt.io diagram of the Microsoft 365 attack chain: an infostealer on the host exfiltrates cookie files, the ESTSAUTHPERSISTENT cookie is replayed through a PKCE request with prompt set to none, a Graph bearer token is issued, and the operator enumerates the tenant, reads mail and performs shell reconnaissance against two victims.
Image: Hunt.io (fair use).How replayed session cookies became Graph tokens, from the Hunt.io report.

Four PowerShell scripts went after 35 AI products, with Claude and Codex session tokens at the top of the list, alongside Gemini sessions, to reach conversation history, project data and organizational API keys. The loot also held 6,570 cryptocurrency credential records from 67 hosts. To keep access after cleanup, one script enrolled machines in ScreenConnect through its API, and another queued a silent Level.io agent install carrying the operator's API key, wired into Phantom53 as a named alias. On one corporate host, reconnaissance showed AnyDesk was already installed.

Hunt.io tied the operation to Russian speakers through Russian prose in comments, docstrings and logs across 249 files, including all 771 lines of the bridge's log. It did not name a known group.

Hunt.io AttackCapture file manager view of the exposed directory on port 8124, listing 1,692 files across nine subdirectories with a total size of 1.11 GB.
Image: Hunt.io (fair use).The open directory as captured by Hunt.io AttackCapture.

What defenders can use

The report gives SOC teams concrete hooks: pythonw.exe running .pyw files from AppData\Roaming or spawned by an unusual parent, uniform POST beacons to workers.dev subdomains, the 12 Workers domains and server addresses in its indicator table, unexpected ScreenConnect or Level.io enrollments, and Microsoft 365 sign-ins that arrive through prompt=none flows from unfamiliar infrastructure. Hunt.io recommends Conditional Access with Continuous Access Evaluation and device-bound sessions, shorter refresh token lifetimes, and treating AI command-line credential files such as ~/.claude/.credentials.json and ~/.codex/auth.json with the same care as SSH private keys.

For banks and agencies, Hunt.io said, the fix sits on the account side: phishing-resistant authentication, session and device binding, and fast resets for affected customers.

About the Author

Priya Shah

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

Related stories

Threat Intel desk