Sysdig Observes Targeted LiteLLM SQL Injection Attempts 36 Hours After Disclosure

Summary
Sysdig TRT observed targeted pre-auth SQL injection attempts against LiteLLM 36 hours after disclosure of CVE-2026-42208, with UNION payloads probing credential tables. No authenticated follow-on or confirmed extraction was observed; Sysdig recommends patching and key-
Key points
- CVE-2026-42208 is a pre-authentication SQL injection affecting LiteLLM versions 1.81.16 through 1.83.6; the Authorization bearer value was concatenated into a PostgreSQL query, allowing arbitrary SELECT statements.
- Sysdig TRT observed the first exploitation attempt on April 26, 2026, 36 hours and seven minutes after the advisory was indexed in the GitHub Advisory Database.
- The observed UNION payloads probed LiteLLM_VerificationToken, litellm_credentials, and litellm_config, which contain virtual keys, provider credentials, and environment configuration.
- Requests came from two egress IPs with the same reported autonomous system and user-agent; Sysdig assessed the activity as consistent with one operator rotating egress IPs.
- Sysdig observed no authenticated calls using extracted keys, virtual-key creation, or confirmed credential reuse; the article does not establish successful extraction or compromise.
- The fix is LiteLLM v1.83.7 or later. Sysdig recommends rotating stored keys and credentials for vulnerable internet-reachable instances, restricting proxy access, and monitoring for the described request patterns.
Article Details
- Attack Vectors
- An operator attempted pre-authentication SQL injection against LiteLLM's proxy authentication path by placing UNION-based payloads in the Authorization: Bearer header.
- The payloads targeted LiteLLM_VerificationToken, litellm_credentials, and litellm_config to enumerate virtual keys, provider credentials, and environment variables; column-count discovery was also attempted.
- The operator repeated payloads from a second egress IP and then probed /key/generate and /key/info without authentication. Sysdig reported no confirmed authenticated follow-on or successful compromise.
- Defensive Notes
- Upgrade LiteLLM to v1.83.7 or later; the article identifies versions 1.81.16 through 1.83.6 as affected.
- Rotate virtual API keys, master keys, and provider credentials for internet-reachable LiteLLM instances running vulnerable versions.
- Restrict proxy access to an internal network or an authenticated reverse proxy. If patching is delayed, block suspicious Authorization header values containing quotes or SQL keywords.
- Monitor web-server logs for Bearer values beginning with sk-litellm' and UNION SELECT payloads targeting the named LiteLLM tables.
- Review upstream provider billing and traffic for requests from unfamiliar IPs following exposure.
Indicators of compromise
| Type | Indicator | Context |
|---|---|---|
| IPV4 | 65[.]111[.]25[.]67 | Second source IP observed replaying SQL injection payloads and probing LiteLLM key-management endpoints without authentication. |
| IPV4 | 65[.]111[.]27[.]132 | Source IP observed issuing UNION-based SQL injection payloads and schema-enumeration probes against LiteLLM. |
MITRE ATT&CK
T1190 · Exploit Public-Facing ApplicationThe operator attempted to exploit a pre-authentication SQL injection in the internet-reachable LiteLLM proxy.T1567 · Exfiltration Over Web ServiceThe UNION payloads attempted to retrieve database values through HTTP responses; the article reports no confirmed successful extraction.
CVE
Vendors
Amazon Web ServicesA single litellm_credentials row often holds an OpenAI organization key with five-figure monthly spend caps, an Anthropic console key with workspace admin rights, and an AWS Bedrock IAM credential.Anthropicin LiteLLM, the open-source LLM gateway with 22,000+ GitHub stars used as a front end for OpenAI, Anthropic, and other model providers.OpenAISQL injection in LiteLLM, the open-source LLM gateway with 22,000+ GitHub stars used as a front end for OpenAI, Anthropic, and other model providers.
Products
Amazon Bedrocklitellm_credentials.credential_values holds the actual upstream provider keys (OpenAI, Anthropic, Bedrock).LiteLLMSysdig TRT observed a rapid, pre-auth SQL injection (CVE-2026-42208 / GHSA-r75f-5x8p-qvmc) against LiteLLM that used UNION-based payloads to enumerate tables holding virtual API keys, provider credentials, and
