CVE-2024-1455: 
Python vulnerability analysis and mitigation

A vulnerability (CVE-2024-1455) was discovered in the langchain-ai/langchain repository, specifically affecting the XMLOutputParser component. The vulnerability allows for a Billion Laughs Attack, a type of XML External Entity (XXE) exploitation, which was disclosed on March 26, 2024. The vulnerability affects users who combine an LLM (or agent) with the XMLOutputParser and expose the component via a web service endpoint (NVD, Huntr).

The vulnerability exists in the XMLOutputParser component which uses the etree module from Python's standard library XML parser. The issue stems from the parser's handling of nested entity references in XML documents. The vulnerability has been assigned a CVSS v3.0 base score of 5.9 (Medium) with the vector string CVSS:3.0/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H. It is classified as CWE-776: Improper Restriction of Recursive Entity References in DTDs (Huntr).

When exploited, the vulnerability can cause the XML parser to consume excessive CPU and memory resources (up to 3GB of RAM or more depending on the number of entities), potentially leading to a denial of service (DoS) condition. This can result in system unresponsiveness or significant performance degradation, affecting the availability of the service to legitimate users (Huntr).

The vulnerability is exploitable when a malicious party attempts to manipulate the LLM to produce a malicious payload for the parser. The attack requires network access but no privileges or user interaction. The attack complexity is considered high according to the CVSS score (NVD, Huntr).

The vulnerability has been fixed by replacing the standard xml.etree.ElementTree with defusedxml.ElementTree in the XMLOutputParser implementation. A patch has been committed to address this issue (Github Commit).