Arize Phoenix Format String Injection in applyChatTemplate (CVE-2026-53998)
Detects exploitation and exposure of CVE-2026-53998, a CWE-134 format string injection in Arize Phoenix's applyChatTemplate functionality affecting arize-phoenix >= 13.0.0 and < 15.7.0 (CVSS 10.0, PoC public). Unsanitized user-controlled input flows into a Python format/template evaluation path, allowing attackers to read in-process memory, leak secrets (CWE-200), or crash the service. This detection surfaces vulnerable Phoenix deployments, suspicious format-specifier payloads sent to Phoenix HTTP endpoints, and Python exceptions/crashes indicative of injection attempts.
Vulnerability Intelligence
Public PoCAffected Software
- Vendor
- pip
- Product
- arize-phoenix
- Versions
- >= 13.0.0, < 15.7.0
Timeline
- Disclosed
- November 17, 2026
References & Proof of Concept
- https://github.com/Arize-ai/phoenix/security/advisories/GHSA-gfv3-m769-5xm3
- https://nvd.nist.gov/vuln/detail/CVE-2026-53998
- https://github.com/Arize-ai/phoenix/pull/13197
- https://github.com/Arize-ai/phoenix/commit/e7224a685589f2ffae1fefcc68ff210bd6ad28b8
- https://github.com/Arize-ai/phoenix/releases/tag/arize-phoenix-v15.7.0
- https://github.com/advisories/GHSA-gfv3-m769-5xm3
CVSS
- Attack Vector
- Network
- Attack Complexity
- Low
- Privileges Required
- None
- User Interaction
- None
- Scope
- Changed
- Confidentiality
- High
- Integrity
- High
- Availability
- Low
What is CVE-2026-53998 Arize Phoenix Format String Injection in applyChatTemplate (CVE-2026-53998)?
Arize Phoenix Format String Injection in applyChatTemplate (CVE-2026-53998) (CVE-2026-53998) maps to the Initial Access and Collection tactics — the adversary is trying to get into your network in MITRE ATT&CK.
This page provides production-ready detection logic for Arize Phoenix Format String Injection in applyChatTemplate (CVE-2026-53998), covering the data sources and telemetry it touches: IIS Logs, Azure App Service HTTP Logs, Network Proxy/WAF. The queries below are rated critical severity at medium confidence, and ship for 7 SIEM platforms — KQL, SPL, Elastic, QRadar, Sumo, YARA-L, LogScale.
MITRE ATT&CK
- Tactic
- Initial Access Collection
// df00tech DF00-b7ac50596e · https://df00tech.com/detections/CVE-2026-53998 · licence: https://df00tech.com/detections/licence
let suspiciousFormat = dynamic(["{0.__class__", "{0.__init__", "{.__globals__", "{0.__class__.__mro__", "%(", "{event.", "{self.", "__builtins__", "__globals__"]);
let phoenixPaths = dynamic(["/v1/chat", "/v1/prompts", "applyChatTemplate", "/graphql", "/v1/datasets"]);
union isfuzzy=true W3CIISLog, AppServiceHTTPLogs, CommonSecurityLog
| extend url = coalesce(column_ifexists("csUriStem", ""), column_ifexists("CsUriStem", ""), column_ifexists("RequestUri", ""), column_ifexists("DestinationServiceName", ""))
| extend query = coalesce(column_ifexists("csUriQuery", ""), column_ifexists("CsUriQuery", ""), column_ifexists("RequestClientApplication", ""))
| extend body = coalesce(column_ifexists("Message", ""), column_ifexists("AdditionalExtensions", ""))
| where phoenixPaths has_any (url) or url has_any (phoenixPaths)
| where url has_any (suspiciousFormat) or query has_any (suspiciousFormat) or body has_any (suspiciousFormat)
| project TimeGenerated, SourceIP = coalesce(column_ifexists("cIP", ""), column_ifexists("CIp", ""), column_ifexists("SourceIP", "")), url, query, body
| order by TimeGenerated desc Identifies HTTP requests to Arize Phoenix endpoints containing Python format-string / introspection payloads (e.g. {0.__class__}, __globals__) characteristic of CVE-2026-53998 exploitation.
Data Sources
Required Tables
False Positives
- Legitimate prompt templates that intentionally use curly-brace placeholders such as {variable} in chat templates
- Security scanners and vulnerability assessment tools probing the Phoenix endpoint
- Developers testing prompt templating locally with literal brace syntax
Query fingerprint DF00-b7ac50596e · licence
Sigma logsource & community rules
The detection logic for Arize Phoenix Format String Injection in applyChatTemplate (CVE-2026-53998) (CVE-2026-53998) above is provided in a vendor-neutral
form so you can deploy it on any SIEM. The same logic is shipped here as native
KQL (Microsoft Sentinel / Defender), SPL (Splunk), Elastic (Elastic Security (EQL)), QRadar (IBM QRadar (AQL)), Sumo (Sumo Logic CSE), YARA-L (Google Chronicle / SecOps), LogScale (CrowdStrike LogScale (CQL)) queries. In Sigma terms, this detection targets the
following logsource:
logsource:
category: network_connection
product: windows Browse the community-maintained Sigma rules for this technique:
Platform-specific guides for CVE-2026-53998
References (6)
- https://github.com/Arize-ai/phoenix/security/advisories/GHSA-gfv3-m769-5xm3
- https://nvd.nist.gov/vuln/detail/CVE-2026-53998
- https://github.com/Arize-ai/phoenix/pull/13197
- https://github.com/Arize-ai/phoenix/commit/e7224a685589f2ffae1fefcc68ff210bd6ad28b8
- https://github.com/Arize-ai/phoenix/releases/tag/arize-phoenix-v15.7.0
- https://github.com/advisories/GHSA-gfv3-m769-5xm3
Testing Methodology
Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.
- Test 1 Phoenix format-string payload via applyChatTemplate endpoint
Expected signal: Web/proxy log entry for POST /v1/chat containing the {0.__class__...__globals__} payload; Phoenix application traceback for a format/template evaluation error.
- Test 2 Phoenix percent-format disclosure probe
Expected signal: Access log entry for POST /v1/prompts carrying the %(__builtins__)s payload.
- Test 3 Vulnerable arize-phoenix version install and enumeration
Expected signal: Process/package telemetry showing pip install of arize-phoenix 15.6.0 (within the vulnerable >=13.0.0,<15.7.0 range).
Response Playbook
Triage
- Confirm whether the targeted host runs arize-phoenix and identify its version: `pip show arize-phoenix` or inspect the container image. Versions >= 13.0.0 and < 15.7.0 are vulnerable.
- Review the flagged request payload for Python format-string / introspection tokens ({0.__class__}, __globals__, __builtins__, __mro__) sent to applyChatTemplate / chat / prompt endpoints.
- Determine whether the Phoenix instance is internet-exposed or reachable by untrusted users, since AV:N/PR:N makes unauthenticated remote exploitation feasible.
- Correlate the source IP against known scanners, CI/CD runners, and legitimate users to separate opportunistic probing from targeted exploitation.
Containment
- Upgrade arize-phoenix to 15.7.0 or later across all deployments; this is the authoritative fix (commit e7224a685589f2ffae1fefcc68ff210bd6ad28b8).
- If immediate patching is not possible, restrict network access to the Phoenix UI/API to trusted networks via firewall/WAF and block requests containing brace-based introspection payloads.
- Rotate any secrets (API keys, DB credentials, JWT secrets) reachable from the Phoenix process, as CWE-200 secret disclosure may have occurred.
Evidence Collection
- Capture full web/proxy/WAF logs for all requests to Phoenix endpoints from the suspect source, including request bodies and query strings.
- Collect Phoenix application logs and Python tracebacks showing format/template exceptions or unexpected object reprs around the event window.
- Preserve the Phoenix process memory/core dumps if a crash occurred, and snapshot the container/host for forensic analysis.
Escalation Criteria
- ! Escalate to incident response if payloads successfully returned object internals, secrets, or environment data in the HTTP response (confirmed data disclosure).
- ! Escalate if the vulnerable version is confirmed on an internet-exposed instance with evidence of repeated exploitation attempts from external sources.
Investigation Guide
Forensic Artifacts
- >
Web server / proxy / WAF access logs containing format-string payloads to Phoenix endpoints - >
Phoenix application logs and Python tracebacks referencing str.format / template evaluation - >
Installed arize-phoenix package version metadata (pip, requirements.txt, container manifests)
Tuning Guidance
Baseline legitimate prompt templates in your environment first, many Phoenix prompts legitimately contain single curly-brace placeholders like {question}. Tighten the regex to dunder/introspection tokens (__class__, __globals__, __builtins__, __mro__) and percent-format sequences rather than any brace to minimize false positives, and allowlist known internal scanner and CI source IPs.
Hunting Queries
Hunts for repeated format-string injection probes against Phoenix chat/prompt endpoints grouped by source IP.
// df00tech DF00-b7ac50596e · https://df00tech.com/detections/CVE-2026-53998 · licence: https://df00tech.com/detections/licence
union isfuzzy=true W3CIISLog, AppServiceHTTPLogs | extend u = coalesce(column_ifexists("csUriStem",""), column_ifexists("CsUriStem","")), q = coalesce(column_ifexists("csUriQuery",""), column_ifexists("CsUriQuery","")) | where u has_any ("applyChatTemplate","/v1/chat","/v1/prompts") | where q has_any ("__globals__","__class__","__builtins__","{self.") | summarize count() by tostring(column_ifexists("cIP","")), u ``` df00tech DF00-b7ac50596e · https://df00tech.com/detections/CVE-2026-53998 · licence: https://df00tech.com/detections/licence ```
index=* (uri="*applyChatTemplate*" OR uri="*/v1/chat*" OR uri="*/v1/prompts*") | regex _raw="__globals__|__class__|__builtins__|\{self\." | stats count by src_ip, uri Atomic Red Team Tests
Sends a crafted chat-template request containing a Python format-string introspection payload to a lab Phoenix instance to trigger CVE-2026-53998.
Command
curl -s -X POST http://127.0.0.1:6006/v1/chat -H 'Content-Type: application/json' -d '{"template":"{0.__class__.__init__.__globals__}","variables":{}}' Cleanup
echo 'No persistent artifacts created; stop/restart the lab Phoenix container if it crashed: docker restart phoenix-lab' Expected Telemetry
Web/proxy log entry for POST /v1/chat containing the {0.__class__...__globals__} payload; Phoenix application traceback for a format/template evaluation error.
Expected Detection
KQL/SPL/EQL rules match the __globals__/__class__ format-string payload sent to the Phoenix chat endpoint.
Sends a percent-style format payload to the prompt endpoint to attempt secret disclosure (CWE-200).
Command
curl -s -X POST http://127.0.0.1:6006/v1/prompts -H 'Content-Type: application/json' -d '{"template":"%(__builtins__)s","variables":{}}' Cleanup
echo 'No cleanup required; no files written.' Expected Telemetry
Access log entry for POST /v1/prompts carrying the %(__builtins__)s payload.
Expected Detection
Detection rules match the %( / __builtins__ format-string markers against the Phoenix prompt endpoint.
Installs a vulnerable Phoenix version in a lab venv to validate version-based exposure detection.
Command
python3 -m venv /tmp/phx && /tmp/phx/bin/pip install 'arize-phoenix==15.6.0' && /tmp/phx/bin/pip show arize-phoenix Cleanup
rm -rf /tmp/phx Expected Telemetry
Process/package telemetry showing pip install of arize-phoenix 15.6.0 (within the vulnerable >=13.0.0,<15.7.0 range).
Expected Detection
Software inventory / vulnerability management flags the installed arize-phoenix version as affected by CVE-2026-53998.
Related Detections
Tactic Hubs
Related Techniques
Same Tactic: Initial Access
Blog posts about this detection
Browse by tactic
- Reconnaissance TA0043
- Resource Development TA0042
- Initial Access TA0001
- Execution TA0002
- Persistence TA0003
- Privilege Escalation TA0004
- Stealth TA0005
- Defense Impairment TA0112
- Credential Access TA0006
- Discovery TA0007
- Lateral Movement TA0008
- Collection TA0009
- Command and Control TA0011
- Exfiltration TA0010
- Impact TA0040