Disclosed weakness · no CVE recorded
LiteLLM could skip token limits on follow-up turns containing Claude tool history
LiteLLM · LiteLLM Proxy · Anthropic Messages token accounting
First documented disclosure · Review checkpoint
Summary
Requests containing Claude server-tool history could skip a configured context limit or reserve too few tokens against the rate limit.
Severity and exploitation evidence
No CVSS score recorded in this review.
Exploitation not established: The reporter observed recurring counting failures. The issue and backport describe no malicious use.
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
The token counter did not recognize some server-tool history blocks. The router continued after the counter failed.
Conditions required
A deployment enforces token-based admission controls and accepts follow-up turns containing Claude server-tool history.
Potential impact
Requests could consume more shared model capacity than their admission checks allowed.
Affected versions and configurations
The reporter observed the behavior in a build identified as 1.110.0 and on main. The precise release range is unspecified.
Fix and mitigation status
LiteLLM lists the fix in its release candidate 1.110.0-rc.3.
Practical next steps
Deploy a build containing the server-tool counting fix and check token-limit accounting for follow-up turns.
Operating lesson
Handle token-counting failures explicitly wherever an estimate determines admission.
Evidence and limits
The maintainer documents a correction. The cited sources provide no issue-specific severity rating or CVE.
Sources
- LiteLLM public issue #51297
Primary source ·
- LiteLLM maintainer fix #51334
Primary source ·
- LiteLLM release-branch backport #51854
Primary source ·
- LiteLLM v1.110.0-rc.3 release notes
Primary source ·