CVE-2026-53998: Critical Arize Phoenix flaw leaks secrets and LLM API keys

A critical format string injection vulnerability in the Phoenix applyChatTemplate GraphQL endpoint allows attackers to extract environment variables, admin credentials, and the master encryption key from the API response.

Christopher Ort
By Christopher Ort · 3 min read
Share
A server room with exposed wiring and open cabinet doors, symbolizing leaked secrets and unauthorized access to critical infrastructure
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. What you should do
  5. What it means
  6. Sources

A critical format string injection vulnerability in the Phoenix observability platform allows attackers to read server-side environment variables, cryptographic secrets, and internal module attributes directly from the API response. The flaw affects the applyChatTemplate GraphQL endpoint and is exploitable without authentication on default installations.

i10X Quick Take

Quick Take

  1. What: Format string injection in Phoenix applyChatTemplate GraphQL endpoint (CVE-2026-53998) exposes environment variables, admin credentials, and the PHOENIX_SECRET master key
  2. Who is affected: arize-phoenix from version 13.0.0 up to, but not including, 15.7.0; default installations with authentication disabled are exploitable without credentials
  3. Severity: CVSS 10.0 Critical; attackers can forge admin tokens and decrypt all stored LLM API keys and OAuth tokens
  4. Fixed in: Version 15.7.0
  5. What to do: Upgrade to arize-phoenix version 15.7.0 or later immediately

In plain English

Phoenix is a tool developers use to monitor AI applications. A security researcher found a serious bug in the part of Phoenix that processes chat templates. An attacker who can reach the Phoenix server over the network can send a specially crafted request that tricks the system into revealing secret information stored on the server, including passwords, encryption keys, and API tokens for AI services. If your organization runs Phoenix to monitor AI systems, you need to upgrade to version 15.7.0 right away.

What you should do: If you don’t use Phoenix, you can ignore this.

What happened

The vulnerability exists in three components of the Phoenix codebase. The applyChatTemplate GraphQL query endpoint (src/phoenix/server/api/queries.py:1638) accepts user-supplied template strings. The FStringTemplateFormatter._format() function (src/phoenix/utilities/template_formatters.py:272) renders these templates without proper sanitization. The DotDict.__getattr__() method (src/phoenix/utilities/template_formatters.py:90) fails to restrict attribute access, allowing traversal to sensitive objects.

According to the advisory, an attacker can send a GraphQL query whose chat template contains a crafted F_STRING placeholder. Because attribute access is not restricted, the placeholder can walk from an ordinary template variable to Python internals and print the full os.environ dictionary, which includes PHOENIX_SECRET, straight into the API response.

On default Phoenix installations, authentication is disabled and the endpoint is reachable without credentials. On deployments with authentication enabled, any account regardless of privilege level can exploit the vulnerability.

The first vulnerable release was arize-phoenix-v13.0.0, published on February 13, 2026. The latest release the reporter confirmed as vulnerable is arize-phoenix-v14.8.0, published on April 17, 2026. The GitHub advisory lists every version from 13.0.0 below 15.7.0 as affected. The issue was reported by the GitHub user OmarXtream.

What you should do

Upgrade to arize-phoenix version 15.7.0 or later immediately. The reporter recommended three combined mitigations: restricting attribute traversal at the Formatter.get_field() level, adding an IsAuthenticated permission to the applyChatTemplate resolver, and limiting the F_STRING format to trusted system-defined templates. Because exposed deployments may already have leaked secrets, teams should also rotate PHOENIX_SECRET and the LLM API keys and OAuth tokens stored in Phoenix after upgrading.

Organizations that cannot upgrade immediately should restrict network access to the Phoenix server and enable authentication if it is currently disabled.

What it means

Successful exploitation allows an attacker to extract PHOENIX_SECRET, the master key used to sign all JWTs and derive the Fernet encryption key protecting stored credentials. With this key, an attacker can forge admin-level session tokens and decrypt all secrets stored in the database, including LLM API keys and OAuth tokens.

The vulnerability received a CVSS score of 10.0, the maximum severity rating. The attack vector is network-based, requires no privileges, no user interaction, and has a scope change that impacts resources beyond the vulnerable component. The confidentiality and integrity impacts are both rated high.

The vulnerability is tracked as CVE-2026-53998 and GHSA-gfv3-m769-5xm3. It is classified as CWE-134 (Use of Externally-Controlled Format String) and CWE-200 (Exposure of Sensitive Information to an Unauthorized Actor).

Phoenix is the latest AI tool this month where a server bug puts stored model provider keys at risk, after DeepAudit exposed LLM API keys and Git tokens and pgEdge RAG Server leaked provider API keys in error messages. New cases are tracked in our AI Security hub.

Sources

  • GitHub Advisory Database and Arize-ai/phoenix security advisory: GHSA-gfv3-m769-5xm3 (CVE-2026-53998), published November 17, 2026
  • Arize Phoenix release notes: arize-phoenix-v15.7.0
Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. CVE-2026-53998: Critical Arize Phoenix flaw leaks secrets and LLM API keys
  2. Genesys Launches Standalone AI Agents and Arena, With Pricing Based on How Hard Each Task Is
  3. Seedance 2.5 4K Reaches Developers: 10-Bit Video up to 30 Seconds With 480p Draft Mode
  4. Perplexity Computer Adds Multiagent Mode, Where Up to 16 Agents on Rival Models Check Each Other
  5. Gemini Can Now Make Calls, Pay Bills and Shop, Free in Six Countries With Gemini 4 Hydrogen