AutoGPT Insecure Deserialization (CVE-2026-33233) Leads to RCE

The National Vulnerability Database (NVD) has reported CVE-2026-33233, a critical insecure deserialization vulnerability affecting AutoGPT versions 0.6.34 through 0.6.51. AutoGPT, a platform for managing continuous AI agents, is susceptible to remote code execution (RCE) due to its Redis cache handling. The backend deserializes Redis cache bytes using pickle.loads without any integrity or authenticity checks.

This means if an attacker can poison a shared-cache key in Redis, they can execute arbitrary commands within the backend container. The NVD highlights that the write path serializes values with pickle.dumps into Redis, and the read path blindly invokes pickle.loads on those bytes. The absence of HMAC/signature validation or strict schema validation on deserialization creates a direct path for compromise, impacting confidentiality, integrity, and availability. The issue has been fixed in version 0.6.52.

Rated with a CVSS score of 7.6 (HIGH), this vulnerability is a prime example of why robust input validation and secure deserialization practices are non-negotiable. Defenders need to recognize that shared cache environments, while efficient, present significant attack surfaces if not properly secured. Blind deserialization is a gift to attackers seeking to achieve RCE.