A critical security flaw, tracked as CVE-2025-48157, has been identified in ChromaDB, a widely used vector database for LLM applications. The vulnerability enables unauthenticated attackers to achieve Remote Code Execution (RCE) and complete server hijacking via exposed FastAPI interfaces.

The exploit targets a failure in how the service processes input, specifically within embedding functions. By injecting malformed numerical data—such as NaN (Not a Number) or Infinity values—into base64-encoded embeddings, attackers can bypass validation and execute arbitrary code on the host system. Given the central role vector databases play in AI pipelines, this flaw allows for full system compromise, unauthorized exfiltration of sensitive vector metadata, and lateral movement into broader AI infrastructure.

ChromaDB has addressed this vulnerability in release 1.5.9, which implements strict rejection of NaN and Infinity values in embeddings. Security teams and CISOs must prioritize the immediate update of all ChromaDB instances to version 1.5.9 to mitigate this risk. Failure to patch exposed instances leaves the entire AI stack vulnerable to total takeover by remote actors.