A format-string flaw in Arize Phoenix's applyChatTemplate query (CVE-2026-53998, CVSS 10.0, disclosed November 16, 2026) lets anyone who can reach a default install read PHOENIX_SECRET and forge admin tokens; fixed in Phoenix 15.7.0 — upgrade and rotate the secret.
What changed
Arize AI's security advisory GHSA-gfv3-m769-5xm3 and the CVE-2026-53998 record, published November 16, 2026, describe a format-string injection (CWE-134) in the applyChatTemplate GraphQL query of Phoenix, Arize's open-source LLM tracing and evaluation platform. On default installations with authentication disabled the endpoint needs no credentials at all, and on authenticated deployments any account, whatever its privilege level, can exploit it. The flaw scores CVSS 10.0 (CVSS:3.1 AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:L).
A successful attack reads the server's environment variables and internal configuration straight from the API response, including PHOENIX_SECRET — the master key Phoenix uses to sign its JWTs and derive the Fernet key that encrypts stored credentials such as LLM API keys and OAuth tokens. With that key an attacker can forge admin-level session tokens and decrypt every secret in the database. The CVE record covers every Phoenix release from 13.0.0 before 15.7.0 (the advisory confirms 13.0.0 through 14.8.0). The fix shipped in Phoenix 15.7.0 on May 13, 2026: upgrade, and rotate PHOENIX_SECRET, because a key that may already have leaked stays valid until it is changed.
How to fix
- Update: Upgrade Phoenix to 15.7.0 or later — the first release that fixes CVE-2026-53998 (fix released May 13, 2026).
- Configuration: Rotate PHOENIX_SECRET after upgrading, and re-issue the credentials Phoenix stores (LLM API keys, OAuth tokens): the secret signs session tokens and derives the key that encrypts them, so a leaked secret stays dangerous until it is replaced.
Metrics
| Metric | CVE-2026-53998 |
|---|---|
| CVSS v3.1 base score | 10.0 |
| Affected | Phoenix 13.0.0 up to but not including 15.7.0 |
| Fixed in | Phoenix 15.7.0 (May 13, 2026) |
| CWE | CWE-134 format-string injection |
| Authentication required | None on default installs |
Why it matters
Phoenix is the observability layer many teams run beside their LLM apps, and it holds the API keys those apps run on — one unauthenticated request can hand an attacker your model keys and admin access to the tracing platform itself. If your business runs Phoenix anywhere reachable, treat this as a same-day fix: move to 15.7.0, rotate PHOENIX_SECRET, and re-issue any credentials Phoenix stored that the leaked key could decrypt.
Sources
- Arize AI — Format String Injection in applyChatTemplate (GHSA-gfv3-m769-5xm3)
- Arize AI — Phoenix v15.7.0 release (the fix)
- CVE Program — CVE-2026-53998 record
Originally published at iacompa.com
Top comments (0)