A severe remote code execution (RCE) vulnerability was found in vLLM, a popular library for Large Language Model (LLM) serving and inference. The vulnerability is known as CVE-2025-29783 and has a CVSS score of 10, the highest severity rating possible. The bug occurs due to vLLM's integration with Mooncake, a feature that is intended for distributed LLM deployments. When Mooncake is on, vLLM exposes an insecure deserialization interface via ZMQ/TCP, which can potentially be used by remote attackers to run arbitrary code on distributed hosts. As vLLM is very popular, with more than 43k GitHub stars, this vulnerability threatens a lot of organizations that use it for LLM deployment. The reason behind this vulnerability is the utilization of pickle.loads(), which is a known security risk when dealing with untrusted data. In particular, vLLM's recv_tensor() method invokes recv_impl, which deserializes raw network bytes directly with pickle.loads(), opening it up to malicious payloads. Attackers can utilize this bug to remotely execute arbitrary code on systems that run vulnerable vLLM versions. The bug affects vLLM versions 0.6.5 up to and including 0.7.x, leaving many deployments vulnerable if not patched. Since distributed machine learning deployments usually consist of several networked servers, this vulnerability greatly expands the attack surface, and thus makes it even more risky. In order to avoid this vulnerability, users should immediately upgrade to vLLM version 0.8.0 since the problem has been resolved in PR #14228. All affected versions currently deployed should apply the update at the earliest opportunity. Also, if the Mooncake integration is not required, the developers should turn it off. Developers must go through their security policies to avoid using unsafe deserialization mechanisms like pickle.loads() on untrusted inputs. Ongoing security auditing and network segregation can also mitigate future risks from such vulnerabilities.
Excessive CPU and memory usage in Google Chrome has become a common issue, often caused by malicious or poorly designed browser extensions. Some extensions are built with hidden mo...
Cybersecurity researchers have identified a new variant of the GlassWorm campaign that targets developers through a malicious Open VSX extension named “specstudio.code-wakatime-a...
A recently disclosed flaw in the widely used Android library EngageSDK has created significant concern within the cryptocurrency space, as it may have exposed millions of users to ...