LLM inference engines vulnerable to classic code injection flaws

A significant security vulnerability, CVE-2025-9141, has been identified in inference engines like vLLM and SGLang, stemming from the use of the `eval()` function on model-generated arguments. This issue, reminiscent of older vulnerabilities like SQL injection and insecure deserialization, arises from treating LLM output with unearned trust rather than as untrusted text. The vulnerability highlights a broader trend of rapid development in AI infrastructure, where security best practices are often overlooked in favor of performance benchmarks, leading to potential exploitation through adversarial inputs. AI

IMPACT Highlights the need for robust input validation in AI infrastructure, as classic software vulnerabilities can reappear in LLM applications.

RANK_REASON Identifies a specific vulnerability in AI inference software, but does not represent a new model release or core AI research.

Read on dev.to — LLM tag →

• CVE-2025-9141
• eval()
• Insecure Deserialization
• SGLang
• SQL injection
• vLLM
• XML external entity attacks

• product
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →