{"id":"CVE-2026-61539","published":"2026-08-21T20:37:02Z","modified":"2026-08-24T19:45:13Z","title":"Xinference: Remote code execution via unsafe `eval()` in Llama3 tool-call parsing","vendor":"xorbitsai","product":"inference","tier":"tail","description":{"text":"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.","provenance":"source","from":"CVE.org"},"cvss":{"score":10,"severity":"CRITICAL","version":"3.1","vector":"CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H","provenance":"source","from":"CVE.org CNA"},"epss":{"score":0.01226,"percentile":0.6956,"provenance":"source"},"cwes":[{"id":"CWE-95","name":"Improper Neutralization of Directives in Dynamically Evaluated Code"}],"references":[{"url":"https://github.com/xorbitsai/inference/security/advisories/GHSA-x2rj-828p-hx9m","tags":["x_refsource_CONFIRM"]},{"url":"https://github.com/xorbitsai/inference/pull/4786","tags":["x_refsource_MISC"]},{"url":"https://github.com/xorbitsai/inference/commit/1b3d220f342ce68d34cec4586d9409d457dadc42","tags":["x_refsource_MISC"]},{"url":"https://github.com/xorbitsai/inference/releases/tag/v2.7.0","tags":["x_refsource_MISC"]}],"priority":{"score":20.9,"factors":[{"name":"EPSS percentile","points":13.91},{"name":"CVSS severity","points":7.0}]},"vendorSlug":"xorbitsai","affectedVersions":[{"vendor":"xorbitsai","product":"inference","versions":["< 2.7.0"]}],"ssvc":{"exploitation":"none","automatable":"yes","technicalImpact":"total","timestamp":"2026-08-24T19:45:03.839039Z","from":"CISA ADP","provenance":"source"},"affected":[{"ecosystem":"PyPI","package":"xinference","vulnerableRanges":["<2.7.0"],"patched":["2.7.0"],"from":["GHSA-x2rj-828p-hx9m","PYSEC-2026-3946"],"provenance":"source"}],"summary":{"text":"Xinference (product \"inference\" from xorbitsai) contains an unsafe eval() flaw in its Llama3 tool-call parsing, located in the llama3_tool_parser and LLM utilities components, affecting version 2.5.0 and earlier. An unauthenticated remote attacker who can send requests to the chat completions endpoint with a tools field can craft a prompt that influences the model's tool-call output, causing the server to evaluate attacker-influenced Python expressions and execute commands in the Xinference server process, with no user interaction required.","provenance":"ai","model":"claude-haiku-5-5","generatedAt":"2026-11-20T07:03:34Z"},"euvd":{"id":"EUVD-2026-64146","exploited":false,"provenance":"source"}}