Packages

LiteLLM v1.110.2 and v1.108.13 LTS: Fixing Key Leaks, Duplicate Stream Billing, and GPT-6.2 Caching

LiteLLM shipped coordinated patches on November 19–20, 2026 across v1.110.2 stable, v1.108.13 LTS, and v1.111.0-rc.3. The updates close a /key/info credential leak, stop Responses API streams from billing duplicate calls, resolve worker memory growth, and restore GPT-6.2 prompt caching.

By 0xScorPionSon

8 min read

LiteLLM v1.110.2 and v1.108.13 LTS: Fixing Key Leaks, Duplicate Stream Billing, and GPT-6.2 Caching

LiteLLM v1.110.2 Release architecture showing AI gateway proxy core, cryptographic verification shield, and prompt-cache ring

Executive Summary

  • Featured Package & Ecosystem: litellm (Python / PyPI and Docker AI Gateway), licensed under the MIT License, with 100,732,250 downloads in the trailing 30 days and 3,544,098,888 total downloads on PyPI (Pepy.tech, verified November 21, 2026).
  • Latest Verified Releases (November 19–20, 2026): Stable v1.110.2 (commit c3cafa3, zero database migrations), pre-release v1.111.0-rc.3, v1.109.3, v1.108.13 (LTS), and v1.106.12 on GitHub Releases.
  • Critical Fixes Landed: Closes an authorization flaw on /key/info (#52838, #52925), stops OpenAI Responses API streams from retrying after output reaches the client (#52461, #52616), fixes request-retention memory leaks (#51888, #52600), and restores GPT-6.2 prompt caching (#52906).
  • Supply-Chain & Runtime Evolution: Following two 2026 supply-chain compromises (1.82.7/1.82.8 in March and 1.107.0 on October 21), LiteLLM now signs container images and Helm charts with keyless cosign via Sigstore (LiteLLM Security Incident Report) while migrating core routing and budget enforcement to Rust (v1.112.0-dev.1).

Adoption Telemetry and Release Lineup

LiteLLM serves as the unified LLM proxy and SDK layer for multi-provider AI applications, translating OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure, and Cohere payloads behind a single API. According to Pepy.tech (checked November 21, 2026), litellm has accumulated 3,544,098,888 cumulative PyPI downloads and 100,732,250 downloads over the past 30 days. Independent security telemetry from Datadog Security Labs places its steady-state distribution at roughly 23 million weekly downloads, making its release cadence directly consequential for CI/CD pipelines and AI agent frameworks.

Starting in November 2026, LiteLLM shifted its release schedule from Saturday to Tuesday mornings and launched a one-month Long-Term Support (LTS) pilot on the 1.108 line (LiteLLM Release Cadence Blog). On November 19–20, 2026, the maintainers shipped coordinated updates across five active release branches to address silent prompt-cache regressions, authorization leaks, and memory retention bugs:

Release Line Patched Version Release Date Key Fixes Included Container Tag (ghcr.io/berriai/litellm)
1.111 RC v1.111.0-rc.3 Nov 19, 2026 All 1.110.2 fixes + failed body GC leak (#52600), TLS SSO redirect (#52213), Postgres pool boot (#52378) v1.111.0-rc.3
1.110 Stable v1.110.2 Nov 19–20, 2026 Responses stream retries (#52461, #52616), GPT-6.2 cache (#52906), /key/info leak (#52838), GC leak (#51888) v1.110.2 (c3cafa3)
1.109 Stable v1.109.3 Nov 19, 2026 GPT-6.2 prompt cache (#52906), /key/info leak (#52838), JWT mapping delay (#52350), Wolfi base image (#53008) v1.109.3
1.108 LTS v1.108.13 Nov 19, 2026 Claude Code billing header cache fix (#52922), GPT-6.2 cache (#52906), /key/info leak (#52838), Cloud SQL boot (#52797) lts / v1.108.13
1.106 Stable v1.106.12 Nov 19, 2026 GPT-6.2 prompt cache (#52906), /key/info leak (#52838), updated dnspython/tornado/opentelemetry + Wolfi base v1.106.12
flowchart LR
    A[AI Client / Claude Code] -->|/v1/chat/completions or /v1/messages| B[LiteLLM Proxy v1.110.2]
    B -->|Auth & Key Scope Check #52838| C{Valid Key Owner?}
    C -->|No| D[Reject / Isolate Userless Keys]
    C -->|Yes| E[Rewrite System Prompt as Developer Input #52599]
    E -->|Single Non-Retrying Stream #52461| F[OpenAI GPT-6.2 / Anthropic]
    F -->|Cache Hit + Exact Token Billing| B

Security Patches, Memory Fixes, and Prompt Caching Internals

1. Closing the /key/info Authorization Leak and Auth Delays

Across all five patched lines (v1.106.12, v1.108.13, v1.109.3, v1.110.2, and v1.111.0-rc.3), PRs #52838 and #52925 fix an authorization boundary flaw on the /key/info management endpoint. Previously, a virtual API key that had no associated user (user_id unset) could query /key/info by key hash or alias and read metadata belonging to other userless keys. For multi-tenant proxy deployments issuing service-level or team-level keys without a user binding, upgrading to a patched release closes that cross-key enumeration path.

In addition, PR #52350 fixes a five-minute propagation delay for JWT key mappings without a jwt_issuer when litellm_jwtauth.issuers is configured (a regression present since v1.103.0), and v1.111.0-rc.3 resolves an Admin UI bug (#52213) where SSO logins redirected to http:// when the proxy sat behind TLS termination.

2. GPT-6.2 Prompt Caching and Claude Code Billing Header Fixes

On November 19, 2026, LiteLLM published a post-mortem titled GPT-6.2 Stopped Caching System Prompts Sent Through LiteLLM. Two days earlier, on November 17, OpenAI changed the Responses API so that the top-level instructions field on GPT-6.2 is billed at the full uncached input rate on every call, whereas reusable prompts placed in a developer message at the start of input continue to receive prompt-cache discounts. Because no HTTP errors occurred, monitoring alerts stayed silent while callers on /v1/chat/completions (using openai/responses/, chatgpt/, or Bedrock Mantle) and /v1/messages paid uncached rates for their system prompts.

PR #52599 (backported as #52906) updates LiteLLM to forward system prompts as a developer message at the start of input for GPT-6.2 deployments, requiring zero configuration changes after upgrading. Separately, in v1.108.13 (LTS) and v1.111.0-rc.3, PR #52922 fixes an issue where Claude Code's rotating billing header invalidated prompt caches and inflated input token counts when routed to non-Anthropic models.

3. Stopping Duplicate Billing on Responses API Streams and Worker Memory Leaks

Two high-impact reliability bugs in v1.110.2 directly affected production cost and stability:

  • Duplicate Stream Retries (#52461, #52616): When a Responses API stream failed after partial output or reasoning tokens had already reached the client, LiteLLM retried the request or restarted on a fallback model group—emitting duplicate SSE events to the caller and billing extra upstream API calls. v1.110.2 halts post-output stream retries.
  • Worker Memory Growth (#51888, #52600): A regression introduced in v1.110.0-rc.1 retained every finished request object in memory until a full Python garbage-collection pass (#51888). In v1.111.0-rc.3, PR #52600 also fixes failed calls holding large request bodies in memory, which previously caused proxy workers to balloon to several GiB under load.
  • Claude Sampling & Cost Tracking (#52341, #52313): v1.110.2 stops sending temperature=1 on Claude models that removed sampling parameters across Anthropic, Vertex AI, and Azure AI (preventing HTTP 400 errors), and ensures forwarded Anthropic credentials on /v1/messages honor custom deployment pricing.

4. Supply-Chain Hardening: Keyless Cosign and Curated Checkpoints

As Anaconda's November 19, 2026 security analysis notes, LiteLLM experienced two supply-chain compromises in 2026: 1.82.7/1.82.8 on March 24 (which CloudSEK later estimated exposed over 2,500 organizations and 434,000 CI/CD pipelines) and 1.107.0 on October 21 (where an unrevoked administrative key allowed attackers to poison prebuilt PyPI wheels and Docker tags for roughly two hours).

In response, starting with v1.111.0-dev.3 on November 15, 2026, LiteLLM signs all release container images and Helm charts with keyless cosign backed by Sigstore short-lived certificates and the public Rekor transparency log (LiteLLM Oct 21 Incident Report). Operators can verify every container pull before deploying:

# Verify the LiteLLM v1.110.2 container image signature via Sigstore cosign
cosign verify \
  --certificate-identity https://github.com/BerriAI/project-releaser/.github/workflows/release.yml@refs/heads/main \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com \
  ghcr.io/berriai/litellm:v1.110.2

# Pin exact PyPI version in requirements.txt or pyproject.toml (avoid unpinned upgrades)
pip install --require-hashes -r requirements.txt
# Or run the verified container directly:
docker run -p 4000:4000 \
  -e LITELLM_MASTER_KEY="sk-your-generated-key" \
  -e DATABASE_URL="postgresql://user:pass@host:5432/litellm" \
  docker.litellm.ai/berriai/litellm:1.110.2

Ecosystem Comparison and Safe Upgrade Checklist

Teams deciding how to route multi-provider LLM traffic in Python have several architectural options depending on whether they need a centralized network proxy or an in-process SDK:

Package Latest Verified Version Architecture Role License Key Trade-Offs (Nov 2026)
litellm v1.110.2 / v1.108.13 LTS Centralized AI Gateway Proxy & Python SDK MIT 100+ providers, budget/RBAC controls, migrating hot paths to Rust (v1.112.0-dev.1); requires strict version pinning and cosign checks
pydantic-ai v2.70.0 (Nov 17, 2026) Type-safe Python Agent SDK & Harness MIT Strong static typing, Temporal/DBOS durable execution, patched ReDoS (GHSA-m4cr-m4qc-9hh3) in 0.70.0; in-process only without standalone proxy UI
langchain-litellm 0.11.2 Integration bridge (LangChain ↔ LiteLLM) MIT Connects LangChain runnables to LiteLLM; inherits LiteLLM dependency surface unless pinned to >=1.110.2

Safe Production Upgrade Checklist

  1. Choose the Right Branch: If you prioritize stability over new features, pin v1.108.13 (ghcr.io/berriai/litellm:lts). If you are on 1.110.x, upgrade immediately to v1.110.2—there are no new database migrations between 1.110.1 and 1.110.2 (v1.110.2 Release Notes). Note also that v1.110.2 fixes a startup regression (#52080 in v1.111.0-rc.3) and 1.108.13 fixes Unix-socket Postgres boot failures on Cloud SQL (#52797).
  2. Enforce Image & Wheel Pinning: Never deploy :latest or unpinned pip install litellm in CI/CD pipelines. Always verify ghcr.io/berriai/litellm tags with cosign verify, use pip install --require-hashes, or pull from a curated, build-delayed artifact channel (such as Anaconda's source-built repository, which enforces a build delay and SHA-256 source verification) as recommended by Anaconda.
  3. Prepare for the Rust Core Transition: Looking ahead, v1.112.0-dev.1 on GitHub Releases already introduces Rust-backed budget hold/settle engines admitted at auth (#51535), Rust route-rule and authentication path planning (#51531, #51605), and Rust-only Bedrock message invocation (#51330), alongside Wolfi base image updates on 1.109.3 and 1.106.12 (#53008, #53033). Benchmark staging latency, memory usage, and custom Python callbacks before adopting 1.112.x when it graduates to stable.

Frequently Asked Questions

Does upgrading to LiteLLM v1.110.2 require a database migration?

No. According to the v1.110.2 release notes, there are no new database migrations in v1.110.2 over v1.110.1, making it a low-friction drop-in patch.

Which LiteLLM version fixes the GPT-6.2 prompt caching regression?

The fix (PR #52599 / #52906) is available in v1.111.0-rc.3, v1.110.2 (stable), v1.109.3 (stable), v1.108.13 (LTS), and v1.106.12 (stable). No configuration changes are needed after upgrading.

Why did LiteLLM add keyless Cosign verification?

Following the October 21, 2026 release-pipeline compromise (v1.107.0), LiteLLM eliminated stored signing keys and adopted keyless cosign with Sigstore short-lived certificates and Rekor transparency logging starting in v1.111.0-dev.3 on November 15, 2026.

Who is affected by the /key/info security fix in v1.110.2?

Deployments using virtual API keys that have no associated user (user_id unset) are affected; prior to the fix (#52838, #52925), a userless key could read other userless keys via /key/info by hash or alias.

Should production teams run v1.111.0-rc.3 or v1.110.2?

Run v1.110.2 (or v1.108.13 LTS) in production. Only adopt v1.111.0-rc.3 in staging unless you are specifically hit by the failed-request body GC leak (#52600) or the TLS-terminated SSO http:// redirect bug (#52213).


Sources

More on Packages

Packages

OpenAI @openai/codex 0.171.0: GPT-6.2 Sol Default, TUI Workspace, and Sandbox Hardening

OpenAI released @openai/codex 0.171.0 on November 20, 2026, switching the default CLI model to GPT-6.2 Sol, enabling the TUI workspace with document and image previews by default, and adding remote app-server execution. Backed by 23.2 million weekly npm downloads and a Rust 1.99.1 core, the release also hardens Linux seccomp sandboxing and enterprise MCP sign-in.

8 min read

Packages

FastMCP 4.2.3: Breaking Changes, OAuth & OpenTelemetry Security Fixes, and Production Upgrade Guide

FastMCP 4.2.3 ships with opt-in parent_task_spans tracing and OpenAPI path validation fixes, capping a week of rapid 4.2.x updates that dropped Python 3.10 support and patched critical OAuth and OpenTelemetry ReDoS flaws. With 64.5 million monthly PyPI downloads, upgrading safely requires checking your Python runtime floor, OAuth consent behavior, and deprecated 4.1 skills APIs. Here is the exact changelog breakdown, ecosystem comparison against the official MCP Python SDK 2.3.0, and a step-by-step upgrade checklist.

9 min read

Tell us what is not working

A thirty-minute call, no deck and no obligation. You will leave with a straight answer about whether this is worth building - including if that answer is no.

Start a project

Direct line

hello@codebumble.net
Dhaka, Bangladesh  :   local timeCalls in English or Bangla