Skip to content
← Blog · · df00tech

Phoenix GraphQL Format String Flaw Lets Attackers Steal Server Secrets and Forge Admin Tokens (CVE-2026-53998)

breaking ghsa pip CVE-2026-53998

A newly disclosed GitHub Security Advisory (GHSA-gfv3-m769-5xm3) reports a critical format string injection vulnerability in the applyChatTemplate GraphQL endpoint of Arize Phoenix. According to the advisory, the flaw lives in FStringTemplateFormatter._format() and an attribute-access guard in DotDict.__getattr__() (src/phoenix/utilities/template_formatters.py), reachable via the applyChatTemplate query resolver in src/phoenix/server/api/queries.py. The advisory assigns a CVSS score of 10.0 and notes a public proof-of-concept payload that abuses Python class and attribute traversal (for example __class__.__bases__[0].__init__.__globals__) inside an F-string template variable to break out of the formatter.

Why It Matters

Per the advisory, affected versions span arize-phoenix-v13.0.0 (2026-02-13) through v14.8.0 (2026-04-17). On default installations where authentication is disabled, the endpoint is reachable with no credentials. On auth-enabled deployments, the advisory states any authenticated account, regardless of privilege level, is sufficient. The reported impact is severe. An attacker can read the server's full os.environ, including hardcoded admin credentials in phoenix.auth, and extract PHOENIX_SECRET, the key used to sign JWTs and derive the Fernet key protecting stored LLM API keys and OAuth tokens. The advisory says this enables forging admin session tokens and decrypting all secrets in the database.

What Defenders Should Do Now

  • Inventory any Arize Phoenix deployments and check the running version against the reported affected range (v13.0.0 to v14.8.0).
  • Treat any Phoenix instance reachable from untrusted networks as high risk, especially if authentication is disabled. Restrict network access until a patched release is confirmed and applied.
  • If a patch is not yet available, consider the advisory's suggested mitigations: enforce authentication on the applyChatTemplate resolver, restrict attribute traversal at the formatter's get_field() level, and limit F-string template rendering to trusted, system-defined templates only.
  • As a hunting angle, review GraphQL and API access logs for Phoenix for applyChatTemplate queries containing template variables that reference Python dunder attributes (for example __class__, __globals__, __init__) or os.environ.
  • If compromise is suspected, rotate PHOENIX_SECRET, any LLM API keys or OAuth tokens stored in Phoenix, and invalidate existing sessions and JWTs.

Developing Intel

This is based on a same-day GitHub Security Advisory disclosure. Affected-version boundaries, patch availability, and exploitation status may evolve. We have not independently verified the proof-of-concept. For full technical details, see the original advisory: GHSA-gfv3-m769-5xm3.

Get new detections in your inbox

New ATT&CK coverage plus CISA KEV / CVE detection rules, roughly weekly. No spam, unsubscribe anytime.