CVE-2025-30358: Mesop: class pollution enables DoS and LLM jailbreak

Mesop is a Python-based UI framework that allows users to build web applications. A class pollution vulnerability in Mesop prior to version 0.14.1 allows attackers to overwrite global variables and class attributes in certain Mesop modules during runtime. This vulnerability could directly lead to a denial of service (DoS) attack against the server. Additionally, it could also result in other severe consequences given the application's implementation, such as identity confusion, where an attacker could impersonate an assistant or system role within conversations. This impersonation could potentially enable jailbreak attacks when interacting with large language models (LLMs). Just like the Javascript's prototype pollution, this vulnerability could leave a way for attackers to manipulate the intended data-flow or control-flow of the application at runtime and lead to severe consequences like remote code execution when gadgets are available. Users should upgrade to version 0.14.1 to obtain a fix for the issue.

An adversary with a low-privilege account on a Mesop-based internal LLM chat tool crafts a malicious HTTP request that exploits the class pollution vulnerability. By overwriting Mesop's global conversation state, they inject a forged 'system' role message containing instructions to ignore all previous safety guidelines. The LLM, receiving what it interprets as a legitimate system-level directive, complies — effectively jailbroken without any prompt injection against the LLM itself. The attacker then exfiltrates sensitive data the LLM has access to via RAG or tool calls, or escalates to DoS by corrupting critical runtime state and crashing the server.