CVE-2026-61539
Xinference is an inference API for running open-source, speech, and multimodal models. In 2.5.0 and earlier, Xinference passes attacker-influenced Llama3 tool-call output to eval() in xinference/model/llm/tool_parsers/llama3_tool_parser.py and xinference/model/llm/utils.py. Requests to /v1/chat/completions with a tools field flow through xinference/api/restful_api.py, xinference/model/llm/transformers/core.py, handle_chat_result_non_streaming(), and _post_process_completion() before extract_tool_calls() or _eval_llama3_chat_arguments() evaluates the model-generated Python expression. An unauthenticated remote attacker can influence that output through a crafted prompt and execute commands in the Xinference server process context. This issue is fixed in version 2.7.0.
Scoring
- Severity
- CRITICAL
- CVSS base score
- 10
- CVSS vector
- CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H
- EPSS probability
- 0.66%
- CWE
- CWE-95
- Published
- 2026-08-21
- Last modified
- 2026-08-24
Affected products
- xorbitsai inference
Weakness type
Related vulnerabilities
- CVE-2026-80351 — Apache Camel K: Camel K Tenant repositories reach Maven execution inside operator pod
- CVE-2026-78550 — Improper Input Handling in Okta Access Gateway Management Console Exception Handler
- CVE-2026-48273 — ColdFusion | Improper Neutralization of Directives in Dynamically Evaluated Code ('Eval Injection') (CWE-95)
- CVE-2026-76190 — ColdFusion | Improper Neutralization of Directives in Dynamically Evaluated Code ('Eval Injection') (CWE-95)
- CVE-2026-79678 — Freeipa: idm: freeipa: idp-add eval() reachable before authorization check allows environment disclosure and denial of service
- CVE-2026-85165 — n8n before 2.36.2 Expression Sandbox Bypass via SpreadElement
- CVE-2026-65643 — Eval injection in cPanel 11.138.0.0 and earlier allows remote authenticated users to execute...
- CVE-2026-19295 — Langflow is affected by multiple remote code execution vulnerabilities due to insufficient code-execution policy enforcement