Friday, 20 November 2026
2 agent hacks today▼ 9 vs yesterday (11)

Phoenix flaw lets attackers steal secrets from AI tool

A format string injection in Arize Phoenix exposes environment variables and the master key on default installs with no login required, according to a GitHub advisory.

By The Agentic Times ·

A locked vault's door swings open when someone whispers a hidden password, revealing glowing keys and secret scrolls spilling onto the floor.
A locked vault's door swings open when someone whispers a hidden password, revealing glowing keys and secret scrolls spilling onto the floor. · Illustration: The Agentic Times

A vulnerability in Arize Phoenix, an open-source tool for building and monitoring AI applications, lets attackers read server secrets through a GraphQL endpoint, according to a GitHub security advisory published on 17 November 2026. The flaw, tracked as CVE-2026-53998, affects the applyChatTemplate endpoint. On default Phoenix installations, where authentication is turned off, the endpoint can be reached with no credentials at all.

The advisory describes the issue as a format string injection. A format string is a template with placeholders that get filled in with values. In this case, the advisory says, the way Phoenix renders those templates allows an attacker to traverse internal Python objects and read data they should not be able to see.

According to the advisory, a successful attack discloses server filesystem paths, loaded Python module names, and internal configuration constants. More seriously, an attacker with network access can read the full server environment, including the admin credentials that the advisory says are hardcoded in the phoenix.auth component.

The advisory warns that the attack can extract PHOENIX_SECRET, described as the master key used to sign all JSON Web Tokens and to derive the encryption key that protects stored credentials such as LLM API keys and OAuth tokens. With this key, the advisory says, an attacker can forge admin-level session tokens and decrypt all secrets stored in the database.

The risk is not limited to installations without a login. On deployments where authentication is enabled, the advisory states that any account, regardless of its privilege level, is enough to exploit the flaw. In other words, a low-privilege user could reach admin-level access.

The advisory lists the first vulnerable release as arize-phoenix version 13.0.0, published on 13 February 2026, and the latest confirmed vulnerable release as version 14.8.0, published on 17 April 2026. It names the affected components as the applyChatTemplate GraphQL query resolver, the FStringTemplateFormatter._format() template rendering function, and the DotDict.__getattr__() method, which it describes as having ineffective attribute access control.

A proof of concept is included in the advisory. It shows a GraphQL query that passes a crafted template payload referencing Python internals to reach the server environment variables. The advisory also includes a demonstration image showing the exploit in action.

The advisory recommends three mitigations used together. It suggests restricting attribute traversal at the formatter's get_field() level rather than at the wrapper class, adding an IsAuthenticated permission check to the applyChatTemplate resolver, and limiting the F_STRING format to trusted, system-defined templates only.

The advisory was credited to a researcher who signed the report. It does not state whether a patched release is available, and the structured record describes the status as reported. Operators running affected versions should review the advisory's guidance and check whether their Phoenix instance is reachable without authentication.

Sources

  1. github.comhttps://github.com/advisories/GHSA-gfv3-m769-5xm3