Three LiteLLM Vulnerabilities Could Let Low-Privilege Users Take Over AI Gateways

· Original article ↗

Summary

Obsidian Security disclosed three LiteLLM flaws that can let a low-privilege user gain proxy-admin access and execute code, exposing secrets and enabling response tampering. BerriAI fixed the chain in v1.83.14-stable; the chain was not reported exploited in the wild.

Key points

  • CVE-2026-47101 lets a low-privilege user bypass authorization by setting the allowed_routes field on a virtual API key.
  • CVE-2026-47102 allows a user to change their own role to proxy_admin through the /user/update endpoint.
  • CVE-2026-40217 enables code execution through the Custom Code Guardrail; researchers demonstrated a reverse shell.
  • A successful chain could expose provider keys, stored credentials, prompts and responses, and could enable an attacker to alter model responses in transit.
  • BerriAI fixed the three-flaw chain in LiteLLM v1.83.14-stable; administrators are advised to upgrade, review privileged accounts and callbacks, and rotate exposed secrets if exposure is suspected.
  • Separately, CVE-2026-42271 in LiteLLM MCP preview endpoints was reported exploited in the wild and added to CISA's KEV catalog; it is distinct from this chain.

Article Details

Vulnerability Types
  • Authorization bypass
  • Privilege escalation
  • Sandbox escape leading to server-side code execution
Severity
CVSS 9.9 (Critical) for the full chain, as rated by Obsidian Security.
Exploitation Status
not_reported
Exploit Availability
unknown
Patch Status
available

MITRE ATT&CK

CVE

Vendors

Products

Related Articles