On this page (8 sections)

A critical flaw in Phoenix, the open-source AI observability and evaluation platform from Arize, lets an attacker read server-side environment variables and cryptographic secrets through a single GraphQL query. The issue is tracked as CVE-2026-53998 and is fixed in arize-phoenix 15.7.0, according to the GitHub advisory.

If you run a self-hosted Phoenix instance, especially one reachable from a network you do not fully control, treat this as an urgent upgrade. Because the exposed material includes the key that signs session tokens and protects stored credentials, upgrading alone is not enough: you also need to rotate secrets.

At a glance

What happened

Phoenix exposes a GraphQL query called applyChatTemplate, used to render prompt templates. The advisory, credited to reporter OmarXtream, describes a format string injection in this endpoint. An attacker can supply a template that makes the server evaluate expressions and return the result in the API response, including process environment variables and internal module state.

The GitHub advisory says the endpoint is reachable with no credentials on default installations where authentication is disabled. When authentication is enabled, any account can reach it, regardless of privilege level. The advisory lists the first vulnerable release as arize-phoenix 13.0.0 (2026-02-13) and the latest confirmed vulnerable release as 14.8.0 (2026-04-17); the affected range in the advisory runs up to, but not including, 15.7.0.

Technical details

The weakness classes are CWE-134 (use of an externally controlled format string) and CWE-200 (exposure of sensitive information to an unauthorized actor). In plain terms, Python's F-string style formatting lets a template reference fields on an object, and those references can walk from one object to another. That is normally harmless when the template is written by the developer. Here the template comes from the caller.

According to the CVE description, the applyChatTemplate query passed attacker-controlled F-string templates to the FStringTemplateFormatter rendering function. A helper class meant to restrict what templates could reach did not prevent traversal through Python private attributes. A crafted field could therefore move through object internals and reach the process environment and internal module state. The advisory's suggested fix says attribute traversal should be restricted at the formatter's field lookup rather than in the wrapper class, which points to where the control was ineffective.

The proof-of-concept payload in the advisory is deliberately not reproduced here. What matters for defenders is the outcome, described next.

Why the impact is so high

  • The advisory says an attacker with network access can read the full server environment, including admin credentials hardcoded in the phoenix.auth module.
  • PHOENIX_SECRET is described as the master key used to sign all JWTs and to derive the Fernet encryption key protecting stored credentials such as LLM API keys and OAuth tokens.
  • With that secret, the advisory says an attacker can forge admin-level session tokens and decrypt secrets stored in the database.
  • The CVSS vector marks scope as Changed. The advisory also notes that stored LLM API keys and OAuth tokens can be decrypted, which means the exposure can reach beyond Phoenix itself, though the sources do not explain why that scope value was assigned.

Who is affected

Anyone running the arize-phoenix Python package at version 13.0.0 or later but earlier than 15.7.0. Exposure depends on network reachability:

  • Default installs with authentication disabled: the endpoint needs no credentials, so anyone who can reach the server can query it.
  • Auth-enabled deployments: any valid account, even the lowest privilege, is enough.
  • Instances reachable only on localhost or a private network are less exposed, but any user or compromised host on that network could still reach them.

Teams that put Phoenix on a public hostname for convenience, or share it with many internal users, are the most realistic targets.

How to check your exposure

First, find out which version is installed. This read-only command, run in the Python environment (or container) that runs Phoenix, prints the installed arize-phoenix version. Anything from 13.0.0 up to 15.6.x is vulnerable.

Show the installed arize-phoenix version (run in the Phoenix Python environment)Shell
pip show arize-phoenix

Next, check whether authentication is enabled and who can reach the service. If you can open the Phoenix UI without logging in, authentication is off. Also review reverse proxy, load balancer and firewall rules to see whether the Phoenix port is reachable from the internet or from broad internal ranges.

If you have access logs that cover the GraphQL endpoint, look for unexpected requests mentioning applyChatTemplate from unfamiliar addresses. This grep lists matching lines in a log file; replace the path with your own log location. Absence of hits does not prove safety, since request bodies are often not logged.

Search a log file for applyChatTemplate requests (adjust the path to your logs)Shell
grep -i "applyChatTemplate" /path/to/your/access.log

What to do now

  1. Upgrade arize-phoenix to 15.7.0 or later in every environment, including containers and development machines.
  2. If you cannot upgrade immediately, restrict network access to Phoenix: bind to private addresses or put it behind a VPN or authenticating proxy.
  3. Enable authentication on any instance that runs without it.
  4. Rotate PHOENIX_SECRET after upgrading, so tokens forged with a leaked secret stop working.
  5. Rotate the credentials stored in Phoenix, such as LLM API keys and OAuth tokens, and any other secret present in the server's environment variables.
  6. Change the admin credentials for the instance and review user accounts for any you do not recognize.
  7. Review provider dashboards for unusual API usage on the keys that Phoenix held.

Upgrade the package with pip in the environment that runs Phoenix. This command installs 15.7.0 or newer; then restart the service so the new code is loaded.

Upgrade arize-phoenix to the fixed release (run in the Phoenix Python environment)Shell
pip install --upgrade "arize-phoenix>=15.7.0"
Caution. Restarting Phoenix and rotating PHOENIX_SECRET can log out existing sessions and may affect stored credentials that depend on the old key. Plan a maintenance window and keep a record of the API keys you need to re-enter.
The advisory also suggests hardening ideas for the project: restrict attribute traversal at the formatter's field lookup, require authentication on the applyChatTemplate resolver, and limit F_STRING templates to trusted system-defined ones. These are recommendations to the maintainers; the fix you should apply is the 15.7.0 upgrade. The upstream fix commit and release notes are public.

Detection and monitoring afterwards

If an affected instance was reachable by untrusted parties, assume its secrets may have been read, because the sources give no reliable way to confirm after the fact whether they were. After rotating, watch for admin sessions from new locations, new or changed users, and unexpected usage of the LLM and OAuth credentials Phoenix stored.

Keep Phoenix off the open internet where possible. You can use the free security headers checker and SSL checker on any public hostname fronting it, and Webforta Monitor can alert you to certificate and DNS changes on that hostname. For background on filtering requests to exposed apps, see what is a WAF.

Confidence and open questions

The technical details here come from the GitHub advisory and the CVE record, which are primary sources. Only one independent publisher covers the story so far, so there is little outside confirmation beyond the advisory itself. The sources do not report active exploitation and the CVE is not marked as known exploited in the CISA data provided, but a public proof of concept appears in the advisory, so do not count on that staying true. Dates also differ slightly by source: the advisory was published in the project repository on Nov 9, 2026, by NVD on Nov 16, and added to the GitHub Advisory Database on Nov 17.

Sources

This article is based on the public sources below. Details can change as vendors and researchers publish more; check the original advisories before acting on version numbers.

  1. CVE-2026-53998 - GitHub Advisory Databasegithub.com
  • CVE-2026-53998
  • arize-phoenix
  • format string injection
  • GraphQL
  • secrets exposure