Format String Injection in applyChatTemplate: Arize Phoenix — unauthenticated env/secret disclosure incl. PHOENIX_SECRET (CVE-2026-53998)
Severity: critical · Patch status: patched · Since 2026-11-17 · CVE-2026-53998
A format string injection flaw (CWE-134) in Phoenix's applyChatTemplate GraphQL endpoint lets attackers render arbitrary f-strings via FStringTemplateFormatter, traversing Python object attributes to read os.environ, filesystem paths, module names and internal config from the API response. This exposes PHOENIXSECRET, the master key signing JWTs and deriving the Fernet key protecting stored credentials (LLM API keys, OAuth tokens), enabling forged admin session tokens and decryption of all database secrets. Default installs have auth disabled; on auth-enabled deployments any account suffices. Affects arize-phoenix =13.0.0 (first vulnerable release 2026-02-13) through <15.7.0; fixed in 15.7.0. A public PoC GraphQL payload exists; no in-the-wild exploitation reported. Recommended fixes: restrict attribute traversal at Formatter.getfield(), add IsAuthenticated permission, limit FSTRING to trusted templates.
What to do
- Upgrade arize-phoenix to 15.7.0 or later immediately
- Rotate PHOENIX_SECRET and all credentials/secrets stored in the database (LLM API keys, OAuth tokens), assuming compromise
- Revoke and re-issue JWT sessions/tokens signed with the old secret
- Restrict network access to Phoenix instances; enable authentication if disabled
- Audit logs for suspicious applyChatTemplate queries
Fixed versions
Sources
All live incidents on Threat Watch