CTIPilot
AI-generated · no human review · verify critical claims against the linked source. how it works →
← Back to Daily brief 2026-05-09
HIGHCVE-2026-42208exploitedupdatedNATOA2vulnerability

CVE-2026-42208: LiteLLM Proxy pre-authentication SQL injection in API-key verification, first exploitation attempt 36 hours after the advisory, CISA KEV-listed (CVSS 4.0 9.3)

LiteLLM SQL injection: attackers targeted stored credentials 36 hours after disclosure; patching does not undo exposure

Defender actions

  • On any LiteLLM Proxy that ran 1.81.16 to 1.83.6 reachable from untrusted networks, upgrade to 1.83.7 or later and rotate the upstream provider keys and virtual keys stored in its database.

Analysis

LiteLLM's advisory GHSA-r75f-5x8p-qvmc, published on 2026-04-20, describes a SQL injection in the proxy's API-key verification: the caller-supplied value was not passed to the database as a separate parameter, and the flaw is fixed in 1.83.7 (LiteLLM, 2026-04-20). Bishop Fox confirmed that versions 1.81.16 through 1.83.6 are exploitable without credentials from any LLM API route, through the Authorization: Bearer value reaching PrismaClient.get_data(); the response is always the same HTTP 401, so exploitation runs over a time-based blind channel using PostgreSQL's pg_sleep() (Bishop Fox, 2026-05-06). On a stock Docker Compose deployment with the Postgres sidecar, the application database user is a superuser, which raises the impact beyond reading metadata (Bishop Fox, 2026-05-06). Virtual keys are stored as SHA-256 hashes, but the provider credentials are stored for use (Bishop Fox, 2026-05-06) (Sysdig, 2026-04-27).

Sysdig observed the first exploitation attempt 36 hours and seven minutes after the advisory reached the global database, with queries against LiteLLM_VerificationToken (virtual keys, including the master key), litellm_credentials (stored upstream provider credentials) and litellm_config (Sysdig, 2026-04-27). Sysdig saw no follow-through: no authenticated calls with exfiltrated keys, no virtual-key minting through /key/generate and no reuse of provider credentials, so its finding is the speed and precision of the enumeration rather than a confirmed compromise; it still advises treating any internet-facing instance on a vulnerable version as compromised (Sysdig, 2026-04-27). CISA added the flaw to its Known Exploited Vulnerabilities catalog on 2026-05-08 (CISA KEV catalog, 2026-05-08). The published workaround, general_settings.disable_error_logs: true, closes only one of the two code paths that reach the vulnerable query (Bishop Fox, 2026-05-06).

Cited evidence

The Sysdig Threat Research Team (TRT) observed the first exploitation attempt 36 hours and seven minutes after the advisory was published to the global database.

The novelty of this finding is the speed and precision of the schema-enumeration attempt, not a confirmed compromise.

Sysdig 2026-04-27

Updates2

Correction

Sysdig, not Bishop Fox or LiteLLM, reported the attackers' queries against LiteLLM_VerificationToken, litellm_credentials and litellm_config, and the first attempt came 36 hours and seven minutes after the advisory (Sysdig, 2026-04-27); no source gives 26 hours. The flaw is a SQL injection in the key lookup (LiteLLM, 2026-04-20), and Bishop Fox's analysis is dated 2026-05-06 (Bishop Fox, 2026-05-06).

Improvement

Sysdig's observation was an exploitation attempt without follow-through: it saw no authenticated calls with exfiltrated keys, no virtual-key minting through /key/generate and no reuse of provider credentials, and calls its finding "the speed and precision of the schema-enumeration attempt, not a confirmed compromise" (Sysdig, 2026-04-27). Sysdig still advises treating any internet-facing instance on a vulnerable version as compromised and rotating every virtual key, master key and provider credential it stores (Sysdig, 2026-04-27).

Sources5

Revision history

  1. Published 2026-05-09-migrated
  2. Correction 2026-10-11T1307Z-audit

    Bishop Fox's analysis is dated 2026-05-06, not 2026-04-30, and the three targeted tables come from Sysdig's exploitation observation, which was uncited. Exploitation began about 36 hours after the advisory, not 26; the flaw is a SQL injection, not an authentication bypass; and the evidence record quoted the entry itself. The entry is rewritten with the LiteLLM advisory as primary source, Sysdig's observation, the affected range and the partial workaround, labelled exposure and detection guidance, a rotation action, a rating and an ATT&CK mapping.

    Changed: title headline summary event_date tags cves affected_products techniques sources evidence sourcing_note classification actions body

  3. Improvement 2026-10-21T1149Z-audit

    Sysdig's 36-hour observation is now described as Sysdig describes it, an exploitation attempt with no follow-on use of extracted keys and no confirmed compromise; Sysdig's advice to treat exposed instances as compromised and rotate their keys stands.

    Changed: title summary body evidence headline

PROVENANCE

AI-generated · no human review · this permalink is the shareable record for the finding · verify operationally critical claims against the linked primary source.