CVE-2026-54236 Scanner

vLLM is a product often utilized by artificial intelligence and machine learning developers to efficiently handle large language models. It is employed by organizations seeking scalable solutions for AI-powered applications, offering capabilities that manage computations and storage across distributed systems. The product's flexibility and handling of complex workflows make it a popular choice for research institutions, technology companies, and startups. Users deploy vLLM to align with dynamic AI tasks and initiatives, relying on its architecture to support varying model sizes and operational demands. Major enterprises value vLLM for its adaptability and performance in managing extensive AI data processes.

The vulnerability detected allows remote attackers to potentially leak heap addresses from the product, which can greatly diminish the effectiveness of Address Space Layout Randomization (ASLR). This flaw presents an opportunity for attackers to calculate system memory allocations more accurately, thus posing a risk of further exploits, such as remote code execution. Information disclosure vulnerabilities like this escalate risks by providing adversaries information to craft precise attacks. The problem arises from the Anthropic-compatible router improperly handling errors, specifically within the "/v1/messages" endpoint. Patch oversights are responsible for this continued vulnerability exploitation path in vLLM.

Technically, the vulnerability exists due to how errors are echoed back through the Anthropic-compatible router. Specifically, echoing str(exc) directly in the endpoint "/v1/messages" allows an attacker to infer heap addresses. The issue is exacerbated when the response includes internal error details, such as "_io.BytesIO object at 0x," coupled with status codes indicating a failure. These revelations occur despite attempts to sanitize responses, pointing to an incomplete fix in handling exceptions securely. The problem stems from insufficient validation and sanitization of error messages before they are returned in lightweight API structures.

If malicious actors exploit this vulnerability, they could achieve information disclosure that informs them of memory layout details. Such leaks can lead to precision attacks aiming at remote code execution or privilege escalation within affected systems. The ability to bypass security layers like ASLR provides adversaries with the groundwork for more sophisticated exploits, increasing the risk profile of the exposed service. Over time, consistent exposure to such vulnerabilities may undermine the reliability and security assurances the product seeks to provide its users.

REFERENCES

• https://github.com/advisories/GHSA-hgg8-fqqc-vfmw
• https://github.com/vllm-project/vllm/pull/45119