Disclosed weakness · no CVE recorded

LiteLLM missing guardrail names could silently remove checks

LiteLLM · LiteLLM Proxy · named guardrail configuration

First documented disclosure · Review checkpoint

Summary

Requests could continue without a named guardrail after a typo or removal left the referenced guardrail unregistered.

Severity and exploitation evidence

No CVSS score recorded in this review.

Exploitation not established: Reviewed maintainer reports and controlled checks do not establish malicious exploitation against users.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

Request processing did not fail closed when configured guardrail names had no loaded implementation.

Conditions required

A request, key, team, project, policy or model references a guardrail that is not registered on the serving pod.

Potential impact

Requests proceed without the expected protection. The maintainer demonstrated a deleted keyword-blocking guardrail allowing the formerly blocked output.

Affected versions and configurations

The complete earlier version range is unspecified. Existing behavior remains the default even in the new development release.

Fix and mitigation status

Prerelease 1.112.0-dev.2 adds PR53214’s opt-in general_settings.reject_unregistered_guardrails=true. The option is off by default and is not a complete guardrail coverage fix.

Practical next steps

Enable and verify the opt-in check on a supported build, validate names against the serving pod and test each required route. Treat removed guardrails as policy changes.

Operating lesson

A missing security control must be visible and fail according to an explicit operator policy.

Evidence and limits

The public PR is November 20 and reports controlled real-provider checks. Realtime WebSocket query-string guardrails bypass this check; Responses retrieval attaches post-call pipelines afterward. Only locally loaded controls count, so newly created controls may temporarily return 400. MCP tools/list returns no tools rather than 400, and batch registry changes can drop records. No malicious exploitation or CVSS score was established.

Sources