Phoenix 20.26 blocks compromised LiteLLM and certifi roots
Updated Phoenix
Summary
Phoenix 20.24.0 to 20.26.1 (October 21-25, 2026) exclude the compromised LiteLLM 1.107.0, raise the certifi floor and log browser errors on the server.
- Tool
- Phoenix
- Category
- Agent infra
- Date
- Open source
- No
- License
- Elastic-2.0
- Pricing
- Not stated in the sources we cited.
Arize Phoenix shipped four releases between October 21 and October 25, 2026: 20.24.0, 20.25.0, 20.26.0 and 20.26.1. The reason to upgrade is supply-chain hygiene: Phoenix now refuses the LiteLLM 1.107.0 wheels that shipped a credential stealer and requires a certifi bundle without the compromised emSign roots. The rest is smaller: browser errors reach the server log, and the server's own telemetry carries its version.
Release facts
| Fact | Detail |
|---|---|
| Product | Phoenix, Arize's AI observability and evaluation server |
| Versions | 20.24.0 (Oct 21), 20.25.0 (Oct 22), 20.26.0 (Oct 23), 20.26.1 (Oct 25) |
| License | Elastic-2.0, per PyPI |
| Python | >=3.11, <3.15, per PyPI |
Security and dependency changes (20.24.0)
- LiteLLM 1.107.0 is excluded. Every LiteLLM specifier now carries
!=1.107.0, because, as PR 17087 puts it, the 1.107.0 PyPI wheels "shipped a credential stealer". PyPI has pulled that version; the pin also covers installs of the evals[dev]and[test]extras, which the repository's three-day install cooldown never protected. Phoenix's own lock file stays on LiteLLM 1.105.1. - certifi floor.
arize-phoenixandarize-phoenix-clientnow requirecertifi>=2026.10.18, which drops the emSign roots behind the compromised eMudhra intermediates (PR 17093). The PR notes thatpip install -Uleaves a transitive certifi in place, so without a floor existing environments would keep the old bundle. - Patched Node for contributors. The JavaScript workspace now refuses to run pnpm on Node versions older than the October 2026 security release (24.22.1, or 26.12.1 on the 26.x line). This affects people building Phoenix from source, not the Docker image.
Other changes
- Browser errors in the server log (20.24.0). Uncaught frontend errors are posted to a new internal
POST /client_errorsroute and written as one ERROR record on thephoenix.client_errorslogger, rate-limited per IP (PR 17005). Useful if you run Phoenix for a team and only see server logs. service.versionon server telemetry (20.26.0). The OpenTelemetry resource for Phoenix's own server telemetry now includesservice.version, so self-hosted operators can split traces and logs by release.- Server fixes (20.25.0). A workaround for a graphql-core 3.3.0 subfield cache collision in multi-root queries, and a formatted ID in the experiments 404 message.
- Playground and docs (20.26.0, 20.26.1). MiniMax text model presets were refreshed,
arize-phoenix-evalsmoved to 3.9.3, and new tracing pages cover Pipecat and Microsoft Agent Framework.
Most of 20.25.0 and 20.26.0's feature list is an internal "factory" that turns triaged GitHub issues into pull requests with managed agents. It changes how the Phoenix team works, not what you deploy.
Questions
Do I need to act if I never installed LiteLLM 1.107.0? Upgrading is still the simple fix: it guarantees the bad version can't be resolved and pulls in the certifi floor. If you did install LiteLLM 1.107.0 anywhere, treat the credentials on that machine as exposed.
Does the certifi change break anything? Only connections that relied on the removed emSign roots. The PR notes that certifi 2026.10.19 restores three of the four roots.
Is Phoenix open source? PyPI lists the Elastic License 2.0, a source-available license. The GitHub API does not report a recognized license for the repository.
More on the Phoenix hub, the earlier Phoenix 20.23 launch and the agent infra category. Related observability tools: Langfuse and Opik.
Alternatives
| Tool | License | Pricing note |
|---|---|---|
PhoenixThis launchPhoenix is an Arize-maintained AI observability and evaluation tool with prompt and dataset tooling. | — | Not stated |
LangfuseOpen source platform for tracing, evaluating and improving AI agents and LLM applications. | MIT | Not stated |
OpikOpen-source platform for tracing, evaluating and optimizing LLM applications and agents. | Apache-2.0 | Not stated |
More from Phoenix
Phoenix 21 changes authentication defaults and adds annotation kinds
PhoenixThe server and related clients need a coordinated upgrade, especially for Docker and notebooks.
Category: Agent infra
Pricing: Not stated
Phoenix 20.23 adds OTLP server-log export
PhoenixPhoenix’s October 16 update exports server logs over OTLP and fixes Bedrock tool-result handling in the playground.
Category: Agent infra
Pricing: Not stated
Phoenix v20.21.0 adds Mistral AI cost tracking
PhoenixArize Phoenix v20.21.0, released October 11, 2026, adds Mistral AI model pricing to cost tracking and fixes evaluator metadata loss.
Category: Agent infra
Pricing: Not stated
Questions
What is Phoenix?
The GitHub repository Arize-ai/phoenix is described as "AI Observability & Evaluation". Its docs list features such as LLM-based evaluations, prompt management, a prompt playground, span replay, datasets and experiments.
Is Phoenix open source?
Not stated in the cited sources. The excerpts say the GitHub API does not report a recognized license for the repository, so no license is claimed here.
How is Phoenix used?
According to its docs, you can create datasets by collecting traces or uploading from code/CSV, run experiments to compare different versions of your application, and attach evaluators to datasets so they run automatically during experiments.
Filed on Agent Radar. Claims in this note are limited to the sources above.