Source: Cloud Security Alliance research note on LiteLLM CVE-2026-42208

CSA's analysis of the LiteLLM proxy pre-auth SQL injection: timeline, in-the-wild exploitation about 36 hours after disclosure, what attackers queried, and remediation advice.

Contents

Summary

CSA explains why a SQL injection in an LLM gateway is unusually damaging: the database it reaches holds upstream provider keys, the proxy’s own master and virtual keys, and connection strings. It reconstructs the timeline and recommends rotating every credential on exposed instances.

Key claims

  • CVSS 9.3. The raw bearer token was concatenated into SQL instead of being parameterised; any LLM API route reached it without authentication.
  • Timeline: fix in 1.83.7-stable on 2026-04-19, advisory on 2026-04-20, GitHub Advisory Database indexing 2026-04-24, first exploitation seen 2026-04-26 (per Sysdig), about 36 hours after indexing.
  • Observed payloads were UNION-based and targeted the verification-token, credentials and config tables.
  • Advice: upgrade, rotate virtual keys, master key and provider credentials even without evidence, treat exposed instances since 1.81.16 as possibly compromised; general_settings.disable_error_logs: true is a stopgap only; prefer least-privilege, project-scoped provider keys.

Feeds

LiteLLM

Suggest changes →