Threat reportVulnerabilityTL-2026-2818
GTIG: AI-Era Vulnerability Discovery and Exploitation Surge — In-the-Wild Exploitation of BeyondTrust CVE-2026-1731, LiteLLM CVE-2026-42271 and Langflow CVE-2026-5027
Google Threat Intelligence Group reports disclosures doubled (5,045 in Jan 2026 to 10,477 in Jul 2026) and exploited vulnerabilities rose to 18/month (Jan-Aug 2026) versus 10.5/month in 2025, with 141 distinct exploited vulnerabilities already exceeding 2025's 127. Exploitation is confirmed against AI middleware (LiteLLM CVE-2026-42271, Langflow CVE-2026-5027) and BeyondTrust CVE-2026-1731, which saw multiple threat clusters within days of disclosure deploying SNOWLIGHT, SPARKRAT and cryptominers.
Key facts
| Identifiers | TL-2026-2818, CVE-2026-1731, CVE-2026-42271, CVE-2026-5027, CVE-2025-3248, CVE-2026-48710 |
|---|---|
| Also known as | AI-Era Vulnerability Exploitation Surge |
| Severity | CRITICAL |
| CVSS | 9.9 |
| Exploitation status | Linked CVE exploited in the wild; Listed in CISA KEV: CVE-2026-1731 (added ), CVE-2026-42271 (added ), CVE-2025-3248 (added ), CVE-2026-48710 (added ) |
| EPSS | CVE-2025-3248: 99.9% (99.9th percentile); CVE-2026-42271: 92.5% (99.8th percentile); CVE-2026-1731: 91.7% (99.8th percentile) |
| Affected products | BeyondTrust Remote Support (RS); BeyondTrust Privileged Remote Access (PRA); BerriAI LiteLLM; Langflow Langflow |
| Weaknesses (CWE) | CWE-78, CWE-77, CWE-22, CWE-94 |
| Malware | AgendaCrypt, SNOWLIGHT, SparkRAT, VShell |
| Tools | BeyondTrust, CloudFlare Tunnel, Ligolo-ng |
| MITRE ATT&CK techniques | T1005 Data from Local System, T1053.003 Cron, T1059.004 Unix Shell, T1059.006 Python, T1071.004 DNS, T1078 Valid Accounts, T1098.004 SSH Authorized Keys, T1190 Exploit Public-Facing Application, T1219 Remote Access Tools, T1496 Resource Hijacking, T1505.003 Web Shell, T1552 Unsecured Credentials and 2 more in the attack flow |
| Motivation | FINANCIAL |
| Target sectors | technology, finance, legal, education, retail, health |
| Target regions | North America, Europe, australia |
| Detection rules | 9 (1 shown free) |
| Sources | 11 cited sources |
Detection coverage
As of 2026-09-30, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2818 across Splunk SPL, Microsoft KQL and Sigma, covering 26 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: LiteLLM MCP Test Endpoint Command Injection (CVE-2026-42271) — Exploitation
Detects POST requests to the LiteLLM proxy MCP preview endpoints /mcp-rest/test/connection and /mcp-rest/test/tools/list. These accept a stdio server configuration (command, args, env) that is spawned as a subprocess on the proxy host, giving unauthenticated or low-privilege RCE and theft of upstream LLM API credentials when chained with the Starlette host-header bypass (CVE-2026-48710). CVE-2026-42271 is in CISA KEV and reported as exploited (Qilin, single-source). Legitimate use of these preview endpoints is limited to admin UI sessions on internal networks.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1190, T1059.004
- Data sources: nginx_access_logs, apache_access_logs, aws_elb_access_logs, waf_cef_logs
- Severity: critical
- Confidence: high
Splunk SPL
index=web sourcetype IN ("nginx:plus:access","access_combined","apache:access","aws:elb:accesslogs") earliest=-24h
```CVE-2026-42271: MCP stdio test endpoints spawn attacker-supplied commands on the LiteLLM host```
| eval uri=lower(coalesce(uri_path, uri, url)), method=upper(coalesce(http_method, method))
| where method="POST" AND match(uri, "/mcp-rest/test/(connection|tools/list)")
| fillnull value="unknown" src dest
| stats count AS requests values(uri) AS uris values(status) AS http_status values(http_user_agent) AS user_agents min(_time) AS first_seen max(_time) AS last_seen BY src, dest
| convert ctime(first_seen) ctime(last_seen)
| sort - requestsMicrosoft KQL
// CVE-2026-42271: LiteLLM MCP stdio test endpoints spawn attacker-supplied commands on the proxy host
CommonSecurityLog
| where TimeGenerated > ago(24h)
| where RequestMethod =~ "POST"
| where RequestURL matches regex @"(?i)/mcp-rest/test/(connection|tools/list)"
| project TimeGenerated, DeviceVendor, DeviceProduct, SourceIP, DestinationIP, DestinationPort, RequestMethod, RequestURL, RequestClientApplication, DeviceAction
| summarize Requests=count(), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), URLs=make_set(RequestURL, 10), Actions=make_set(DeviceAction, 5) by SourceIP, DestinationIP, RequestClientApplication
| order by Requests descSigma
title: LiteLLM MCP Test Endpoint Command Injection CVE-2026-42271
id: 3b7e8c1a-5d24-4f0b-9a63-1c2e7f4d8a10
status: experimental
description: Detects POST requests to the LiteLLM proxy MCP preview endpoints /mcp-rest/test/connection and /mcp-rest/test/tools/list, which spawn an attacker-supplied stdio command on the proxy host (CVE-2026-42271, CISA KEV).
references:
- https://www.wiz.io/vulnerability-database/cve/cve-2026-42271
- https://cybelangel.com/blog/itellm-vulnerability-cve-2026-42271/
- https://cloud.google.com/blog/topics/threat-intelligence/vulnerability-discovery-and-exploitation-trends-in-the-ai-era/
author: Threadlinqs Team
date: 2026-09-30
tags:
- attack.initial-access
- attack.t1190
- attack.execution
- attack.t1059.004
- cve.2026-42271
- cve.2026-48710
logsource:
category: webserver
detection:
selection:
cs-method: 'POST'
cs-uri-stem|contains:
- '/mcp-rest/test/connection'
- '/mcp-rest/test/tools/list'
condition: selection
falsepositives:
- LiteLLM administrators testing MCP server connectivity from the admin UI
level: highKnown false positives:
- LiteLLM administrators validating MCP server connectivity from the admin UI or CI smoke tests
Tuning notes: Allow-list internal admin source ranges and known CI hosts. Any hit from an internet source is high fidelity. Pivot on the source IP to process telemetry for shells spawned by the litellm process (see SPL-002). Requires reverse-proxy/WAF logs because LiteLLM itself does not log the request body.
- BeyondTrust thin-scc-wrapper Arithmetic Command Injection (CVE-2026-1731) — Exploitation · unlock with TL Creds
- Cron, SSH authorized_keys and Web Shell Persistence Drops — Post-Exploitation · unlock with TL Creds
- OAST Callback and DNS Exfiltration Domains — C2 / Exfiltration · unlock with TL Creds
- Langflow File Upload Path Traversal and validate/code Injection (CVE-2026-5027 / CVE-2025-3248) — Exploitation · unlock with TL Creds
- Credential and Database Harvesting on Compromised AI Gateway / Appliance Host — Post-Exploitation · unlock with TL Creds
- Tunnelling, Remote-Access and Payload Tooling Post-CVE-2026-1731 — C2 / Lateral Movement · unlock with TL Creds
- Shell, Downloader or Inline Python Spawned by LiteLLM/Langflow — Post-Exploitation · unlock with TL Creds
- CVE-2026-1731 Exploitation Infrastructure and Reverse-Shell Ports — C2 / Ingress · unlock with TL Creds
Indicators of compromise
As of 2026-09-30, Threadlinqs Intelligence tracks 26 indicators of compromise for TL-2026-2818.
Sample indicators (3 of 26)
- IP address:
45.61.150[.]96 - Domain:
aliyundunupdate[.]xyz - URL:
hxxps://64[.]95[.]10[.]115:23011/update.sh
Defanged, TLP:CLEAR. Full set of 26 indicators: unlock with TL Creds.
Technical analysis
Google Threat Intelligence Group (GTIG) published 'Vulnerability Discovery and Exploitation Trends in the AI Era' on 2026-09-30. Monthly CVE disclosures roughly doubled from 5,045 (Jan 2026) to 10,477 (Jul 2026), peaking at 10,740 in Aug 2026. High-risk disclosures rose from 131 to 350 (+167%). Exploitation averaged 18 vulnerabilities/month in Jan-Aug 2026 (10.5/month in 2025); 141 distinct vulnerabilities were exploited in eight months versus 127 in all of 2025, only about 0.23% of disclosures (1 in 431). Zero-days averaged 11/month (8/month in 2025), spiked to 22 in Aug 2026 and made up 62% of exploited vulnerabilities. Edge and security appliances accounted for 14% of exploited flaws (65% of them High/Critical threat risk) because unauthenticated management interfaces sit outside EDR visibility; enterprise directory and collaboration products accounted for 11%. Mass-disclosure events (TOTOLINK router firmware, Apr-May 2026; Oracle and Linux kernel drivers, Jun-Aug 2026) inflated volume.
GTIG reports that AI-discovered vulnerabilities skew toward more consequential classes: 50% lead to remote code execution versus 26% for non-AI-discovered bugs. CVE-2026-1731 (BeyondTrust Remote Support / Privileged Remote Access, unauthenticated OS command injection, discovered autonomously by Hacktron AI) was first exploited about 4 days after disclosure and by 5 additional threat clusters within 7 days, deploying SNOWLIGHT, SPARKRAT and cryptominers, with privilege escalation, lateral movement and data exfiltration. Unit 42 independently documented exploitation via the WebSocket-exposed thin-scc-wrapper script (bash arithmetic-context injection in the remoteVersion parameter), leading to SparkRAT, VShell, web shells (including a China Chopper/AntSword-style aws.php), temporary admin account takeover, SimpleHelp/AnyDesk, Cloudflare Tunnel, Ligolo-ng, Meterpreter, Nezha agent, OAST-based DNS exfiltration and PostgreSQL dumps, across financial, legal, technology, higher-education, retail and healthcare victims in the US, France, Germany, Australia and Canada. BeyondTrust cloud customers were patched 2026-02-02; the advisory and self-hosted patches followed 2026-02-06; CISA added the CVE to KEV on 2026-02-13.
GTIG also tracks AI middleware exploitation. CVE-2026-42271 (BerriAI LiteLLM 1.74.2 through 1.83.6, patched in 1.83.7) is command injection in the MCP preview endpoints POST /mcp-rest/test/connection and /mcp-rest/test/tools/list, where a stdio server configuration (command, args, env) is spawned as a subprocess on the proxy host, yielding host takeover and theft of upstream API credentials; CybelAngel reports it chains with a Starlette host-header validation bypass (CVE-2026-48710) for unauthenticated RCE (CVSS 10.0), was added to CISA KEV on 2026-06-09, and names the Qilin ransomware group as exploiting it (single-source claim, not stated in the GTIG article). CVE-2026-5027 (Langflow <= 1.8.4) is a path traversal in the POST /api/v2/files upload handler where the unsanitized multipart 'filename' allows arbitrary file writes (cron jobs, SSH authorized_keys, webshells); with Langflow's default auto-login it needs no credentials. It was exploited in the wild from June 2026; roughly 7,000 instances are internet-exposed (Censys); fixed in 1.9.0 and later (1.10.0 current as of 2026-06-10). The older Langflow CVE-2025-3248 (unauthenticated Python code injection via /api/v1/validate/code exec()) was added to CISA KEV in May 2025.
GTIG counts 2,076 cumulative AI-stack vulnerabilities (1,500+ in 2026), with agent orchestration frameworks (Flowise, Langflow, LangChain, Dify and others) at 782 in 2026, about half of all AI-related flaws and a +347% increase. GTIG also cites the first known case of a threat actor holding a zero-day exploit script developed with generative AI, intercepted during operational planning (from GTIG's May AI Threat Tracker). No IOCs are in the GTIG article; the IOCs below come from Unit 42's BeyondTrust analysis. The article gives no CVSS scores; scores here come from vendor/third-party sources.
MITRE ATT&CK techniques by tactic:
Collection
Persistence
T1053.003 Cron; T1098.004 SSH Authorized Keys; T1505.003 Web Shell
Execution
T1059.004 Unix Shell; T1059.006 Python
Command and Control
T1071.004 DNS; T1219 Remote Access Tools; T1572 Protocol Tunneling
Defense Evasion
Initial Access
T1190 Exploit Public-Facing Application
Impact
Credential Access
Resource Development
Remediation
Patches
- BeyondTrust BT26-02-RS / BT26-02-PRA (RS 25.3.2+, PRA 25.1.1+)
- LiteLLM 1.83.7 and Starlette 1.0.1
- Langflow 1.9.0 / 1.10.0 (langflow-base 0.8.3)
Immediate actions
- Patch self-hosted BeyondTrust Remote Support to 25.3.2+ and Privileged Remote Access to 25.1.1+ (BT26-02 advisory) and hunt for web shells, new admin accounts and remote-access tools
- Upgrade LiteLLM to 1.83.7+ and Starlette to 1.0.1+; block or restrict /mcp-rest/test/* endpoints; rotate all upstream LLM API keys and stored credentials held by the gateway
- Upgrade Langflow to 1.10.0+ (fix in 1.9.0); disable default auto-login and remove internet exposure of Langflow instances
- Audit process trees on proxy/orchestration hosts for shells, interpreters and network tools spawned by litellm/langflow processes; check crontabs and ~/.ssh/authorized_keys for unauthorized writes
Workarounds
- Restrict BeyondTrust management interfaces and WebSocket endpoints to trusted networks
- Block egress DNS to OAST domains and monitor Cloudflare Tunnel, SimpleHelp and AnyDesk installations on appliance hosts
Longer-term hardening
- Move from volume patching to threat-intelligence-driven triage prioritizing internet-facing edge appliances and KEV-listed flaws
- Sandbox and network-isolate autonomous agent and AI orchestration workloads; never expose AI gateways or workflow builders to the internet unauthenticated
- Compress n-day patch SLAs for edge and AI middleware to days, given exploitation within about 4 days of disclosure
- Deploy pre-release AI-assisted code review to find flaws before production shipping
Affected and fixed versions
- BeyondTrust — Remote Support (RS)
Vulnerable versions: 25.3.1 and prior
Fixed in: 25.3.2 and later (patch BT26-02-RS) - BeyondTrust — Privileged Remote Access (PRA)
Vulnerable versions: 24.3.4 and prior
Fixed in: 25.1.1 and later (patch BT26-02-PRA) - BerriAI — LiteLLM
Vulnerable versions: 1.74.2 through 1.83.6
Fixed in: 1.83.7 and later - Langflow — Langflow
Vulnerable versions: 1.8.4 and prior (CVE-2026-5027)
Fixed in: 1.9.0 and later; 1.10.0
Timeline
- Langflow CVE-2025-3248 (unauthenticated Python code injection via /api/v1/validate/code) added to the CISA Known Exploited Vulnerabilities catalog (May 2025).
- BeyondTrust automatically patches SaaS/cloud customers for CVE-2026-1731.
- BeyondTrust releases advisory BT26-02 and fixes for CVE-2026-1731 (unauthenticated OS command injection in Remote Support and Privileged Remote Access, CVSS v4 9.9).
- CISA adds CVE-2026-1731 to KEV; GTIG reports first exploitation about 4 days after disclosure and 5 additional threat clusters within 7 days.
- Unit 42 publishes analysis of CVE-2026-1731 exploitation: SparkRAT, VShell, web shells, admin account takeover and OAST DNS exfiltration.
- Langflow CVE-2026-5027 path traversal in POST /api/v2/files publicly disclosed (found by Tenable); langflow-base 0.8.3 released 2026-03-30.
- LiteLLM 1.83.7 released fixing CVE-2026-42271 (MCP test endpoint command injection); CVE published the same day.
- CVE-2026-42271 added to CISA KEV after a chained unauthenticated exploit with Starlette CVE-2026-48710 was confirmed; Qilin reported as exploiting it.
- Langflow 1.10.0 released; CVE-2026-5027 exploitation observed in the wild (honeypots see file drops), about 7,000 instances exposed.
- GTIG publishes 'Vulnerability Discovery and Exploitation Trends in the AI Era': 141 exploited vulnerabilities in Jan-Aug 2026 versus 127 in all of 2025; 2,076 cumulative AI-stack CVEs.
Sources cited
- CISA Known Exploited Vulnerabilities Catalog CISA · Government advisory · 2026-02-13
- Vulnerability Discovery and Exploitation Trends in the AI Era (GTIG) Google Cloud · Research · 2026-09-30
- Unit 42: VShell and SparkRAT Observed in Exploitation of BeyondTrust Critical Vulnerability (CVE-2026-1731) Unit 42 · Research · 2026-02-19
- BeyondTrust Security Advisory BT26-02 beyondtrust.com · Vendor report · 2026-02-06
- Horizon3.ai: CVE-2026-1731 horizon3.ai · Research · 2026-02-10
- Wiz Vulnerability Database: CVE-2026-42271 Wiz · Research · 2026-05-08
- Cloud Security Alliance: CVE-2026-5027 Langflow Path Traversal to Unauthenticated RCE labs.cloudsecurityalliance.org · Research · 2026-06-10
- Arctic Wolf: CVE-2026-1731 Unauthenticated OS Command Injection in BeyondTrust RS and PRA arcticwolf.com · Blog · 2026-02-09
- CybelAngel: Top 7 Things to Know About the LiteLLM CVE-2026-42271 Exploit cybelangel.com · Blog · 2026-06-09
- BleepingComputer: Path traversal flaw in AI dev platform Langflow exploited in attacks BleepingComputer · News · 2026-06-10
- The Hacker News: Unpatched Langflow flaw CVE-2026-5027 The Hacker News · News · 2026-06-10
Other reports on CVE-2026-1731
- Storm-1175 Medusa Ransomware Zero-Day Exploitation Campaign (CVE-2026-23760, CVE-2025-10035)
- Seedworm (MuddyWater) Iranian MOIS APT Campaign Targeting U.S. Critical Infrastructure with Dindoor and Fakeset Backdoors
- BeyondTrust Pre-Auth RCE (CVE-2026-1731) — CVSS 9.9, CISA KEV, WebSocket Command Injection, VShell/SparkRAT, 16K+ Exposed Instances, Multi-Sector Campaign
- Ransomware Threat Landscape 2025-2027 — RaaS Destabilization, Conti Leak Cascade, ESXi Hypervisor Targeting, BYOVD as Standard Prep, Double/Triple Extortion Economics, IAB Vertical Integration, Critical Infrastructure Escalation
Other reports on CVE-2026-42271
- PoeLLM (Canto Incognito) malware infects exposed AI servers (LiteLLM, Ollama, Gotenberg, Gitea, Ivanti Sentry) in cryptomining attacks - CVE-2026-42271, CVE-2026-48710
- Qilin-Linked Campaign Exploits MCP Gateway and LLM Framework Flaws (CVE-2026-59822, CVE-2026-42271, CVE-2026-48710) for RCE and Cryptomining
- LiteLLM Proxy Authentication Bypass via Host Header Injection (CVE-2026-49468)
- CVE-2026-42271: LiteLLM MCP Server Command Injection Under Active Exploitation, Chained with CVE-2026-48710 (Starlette BadHost) for Unauthenticated RCE
Previous and next in vulnerability
Further reading
Drafted by the Threadlinqs research pipeline from 11 sources; published ; last updated —.