Skip to content

BREEZE COMET (GTIG)

Tags

Summary

Google Threat Intelligence Group published its first full write-up (September 1, 2026) of BREEZE COMET (formerly UNC5669), a financially motivated threat actor that since 2024 has compromised Brazilian financial-services, retail, and eCommerce organizations to execute fraudulent transfers through Brazil's core payment rails, Pix, STR (Reserves Transfer System), and Boleto, rather than the client-side retail fraud that has historically defined the region's cybercrime ecosystem. GTIG's durable read: a shift from opportunistic banking fraud to direct intrusion into the financial switch itself, supported by a custom multi-language backdoor suite and LLM-generated operational scripts. The activity overlaps with reporting publicly known as Plump Spider and Trend Micro's SHADOW-AETHER-064, the same Brazilian-financial cluster this wiki already tracks via Unit 42's CL-CRI-1163 reporting, making GTIG's tool taxonomy the missing naming layer on an on-wiki campaign.

The objective defines the access requirements, and they are unusually specific: BREEZE COMET needs network access to Brazil's financial system network (RSFN) through an entity that holds it, mTLS credentials that authenticate transactional payloads in the name of an org with funds, persistent multi-account AD/cloud access, and knowledge of the victim's transfer procedures and anti-fraud controls. Everything below serves those four requirements.

Confirmed tradecraft (GTIG/Mandiant incident investigations since 2024)

  • Initial access, several ways in:
  • Early: password spraying and IT-support vishing to talk users into installing RMM tools (AnyDesk observed). Axur corroborates the vishing and adds reported insider recruitment attempts at target organizations.
  • Mid-2025: compromised Brazilian small-government (.gov) websites used both as social-engineering staging (infostealers disguised as tax/receipt documents, e.g. ComprovantePDF.exe, and purchased/cracked XWORM backdoors persisting via startup shortcut edits) and as C2 endpoints, trusted-domain infrastructure that sails past reputation filters. GTIG observed the same staging behavior replicated on municipal domains in Nigeria, Paraguay, Ghana, and Venezuela, and the same staging infrastructure reused across operations targeting multiple organizations.
  • 2025: physically plugging rogue hardware devices into retail store networks to obtain a foothold, then moving laterally and pulling post-exploitation frameworks from external open directories with Netcat plus custom scripts. Trend Micro separately reported JBoss AS exploitation for initial access (the same entry point as SHADOW-AETHER-064).
  • Reconnaissance / AD operations: Impacket, ADRecon, ADVipscan downloaded from GitHub and executed in-memory via PowerShell; a custom LDAP brute-forcing utility, REALBREEZE. Environments with low observability are explicitly profitable.
  • Credential hunting tuned to payments: custom scripts search files and environment variables for mTLS credentials and admin certificates using grep-style terms boleto, cnab, remessa, webhook.*pix, instant.*payment; CI/CD environments are mined for hard-coded pipeline credentials, API keys, and privileged cloud tokens. Lateral movement via hijacked service accounts initiating RDP, SMB exec, and deliberate SMB-path enumeration scans across internal subnets.
  • Segmentation bypass, COBALTSPIN: a Rust lightweight network tunneler that establishes a reverse SOCKS5 proxy over WebSocket back to C2, routing traffic between the operator and internal financial-API infrastructure straight through boundary firewalls with no host persistence mechanism to alert on.
  • Redundant persistence suite (2025-2026, four named backdoors):
  • LIGHTPAINT (Java): installs a legitimate VPN (e.g. SoftEther) configured for automated persistence, then programmatically adds Windows Defender Firewall inbound allow rules for the VPN manager and clears the Windows Networking Vpn Plugin Platform event log to erase the connection's forensics.
  • MILDFROST (Java): passive JAR backdoor living inside the JVM process space; classes such as DnsCommandBeacon.class run slow covert DNS tunnels, and it serves as fallback C2, querying delegated subdomains for instructions and fresh copies of the C++ payloads.
  • KICKPLATE (Nim): payload-delivery and host-persistence workhorse impersonating Windows Update Health Tools; controls the SOCKS5 tunnelers, updates registry startup keys, silently modifies Windows services; supplemented by schtasks.exe as SYSTEM and .lnk edits in startup folders.
  • BOATBEAM (Go): stands up a fake IIS HTTPS server on 443, activating backdoor behavior only when it receives a specific session cookie, otherwise it is indistinguishable from a web server.
  • Defense impairment: direct PowerShell Set-MpPreference -DisableRealtimeMonitoring $true across compromised hosts; post-theft event-log clearing across hosts and deletion of attacker-created directories.
  • Cloud/Kubernetes persistence (2025): malicious Kubernetes pods deployed to maintain access and steal cloud secrets, exfiltrated to public paste/notepad sites (dontpad[.]com).
  • AI usage: Mandiant recovered scripts showing LLM-accelerated development of recon, credential-validation, mass-deployment, victim-specific pivoting, and data-extraction tooling, highly functional code lacking human idiosyncrasies: unrolled structures, verbose explanatory comments, standardized execution headers (GTIG publishes an excerpt with Portuguese comments and per-step banners: ### STEP 1: ENUM ALL LINUX (SSH PORT 22) ###).
  • Payoff: within 24-48 hours of reaching core financial applications via COBALTSPIN + privileged accounts, the actor executed two waves of hundreds of fraudulent transactions (per a client and third-party forensic analysis); at least one confirmed heist of tens of thousands of USD since 2024, modest for the access achieved, which itself is a signal of an actor building capability faster than monetization.

Why it matters (durable reads)

  1. The payment switch is a network target, not an app target. The four access requirements (RSFN reach, mTLS transaction certs, persistent AD/cloud accounts, procedural knowledge) mean detections that work against bank fraud (card/endpoint telemetry) are blind here; the indicators to hunt are mTLS keystore/certificate access with payment-vocabulary search terms, service-account RDP, and SMB-enumeration scans.
  2. The tool names are now durable pivots. COBALTSPIN (Rust reverse-SOCKS5-over-WebSocket), LIGHTPAINT (VPN + firewall-rule + VPN-event-log-clearing trio), MILDFROST (in-JVM DNS-tunnel fallback C2), KICKPLATE (Nim, Windows-Update-Health-Tools masquerade), BOATBEAM (cookie-gated fake-IIS), REALBREEZE (LDAP bruter). Any one of these appearing outside Brazilian finance would mark diffusion of this playbook.
  3. Compromised trusted websites as staging AND C2, replicated across continents. Brazilian municipal .gov sites plus the Nigeria / Paraguay / Ghana / Venezuela municipal replication is the most transferable finding: local-government CMS compromise is becoming commodity crime infrastructure, and domain-reputation controls are exactly the defense it defeats, GTIG's own mitigation guidance says stop relying on .gov allowlists and do TLS decryption + DPI on egress.
  4. Rogue hardware on retail branch networks is the low-tech foothold most organizations don't model; the fix GTIG prescribes (802.1X NAC on switch ports, port security, locked closets) is physical-network work, not endpoint work.
  5. Corroboration, not new attribution. BREEZE COMET (GTIG) overlaps Plump Spider and SHADOW-AETHER-064 (Trend Micro) and Unit 42's CL-CRI-1163; SockTz-era infrastructure (167.148.195[.]53) and JBoss entry previously tied on-wiki all sit in this same space. Treat the names as vendor labels over a likely-porous Brazilian financial-cybercrime ecosystem; no single operator identity is established by any of them.
  6. LLM-generated ops scripts with Portuguese comments are a hunting texture (already our standing rule: AI-style tells only in combination with behavior, here: tunneling, credential search for payment terms, log clearing).

Hunt guidance

  • Firewalls/VPN: alert on inbound firewall rule additions by non-admin/service contexts, VPN service installs (SoftEther especially) outside IT-managed hosts, and clearing of Microsoft-Windows-Networking/VpnPlugin-Platform event logs (log-clear on one host is rare; this specific channel is the tell).
  • JVM estates: passive JARs loaded into long-running Java processes with DNS-querying behavior (DnsCommandBeacon pattern); Java processes making DNS TXT/subdomain lookups they've never made.
  • HTTPS listeners: web servers on 443 not in inventory, or IIS-looking services whose responses differ only when a specific cookie is present (compare cookie/no-cookie fetches from your own scanning).
  • Egress: reverse-SOCKS5/WebSocket tunneling from hosts that have no business tunneling; outbound to paste/notepad services (dontpad[.]com and equivalents) from Kubernetes workloads.
  • Credential stores: file/env searches combining boleto|cnab|remessa|webhook.*pix|instant.*payment, this vocabulary in PowerShell/bash history is close to a one-signature detection for this actor class.
  • Physical: 802.1X enforcement reports and switch-port security logs on retail/branch VLANs; unknown MACs obtaining DHCP.
  • Municipal-site compromise (if you run a local-government CMS anywhere): unexpected file writes, and the fact that your domain appearing in someone else's TLS certificates is a compromise indicator.

Monitoring

  • GTIG follow-ups naming additional victims or the mTLS-credential theft outcomes; any disruption/takedown touching the Nigerian/Paraguayan/Ghanaian/Venezuelan municipal staging set.
  • Whether COBALTSPIN/LIGHTPAINT/MILDFROST/KICKPLATE/BOATBEAM names appear in other clusters (playbook diffusion beyond Brazil).
  • Convergence or divergence with Trend Micro SHADOW-AETHER-064 and Unit 42 CL-CRI-1163 follow-on reporting (167.148.195[.]53 rotation, SockTz versions), does anyone tie the three labels to one operator?
  • XWORM sales/leak-tracking for the cracked versions this actor used, and Axur's insider-recruitment claims maturing into named cases.
  • Expansion GTIG flags as possible: Africa and other Latin American countries, watch Pix-style instant-payment rails elsewhere (Mexico's DiMo SPEI, Colombia's Bre-B) for the same playbook.

Sources

  • Google Threat Intelligence Group / Mandiant: Financially Motivated Threat Actor BREEZE COMET Targets Brazil (September 1, 2026; captured by this wiki September 20, 2026 sweep).
  • Corroborating labels referenced by the primary and our prior pages: Trend Micro SHADOW-AETHER ("Vibe Hacking" report); Unit 42 CL-CRI-1131/CL-CRI-1163; Axur (vishing + insider recruitment); Trend Micro JBoss exploitation reporting.