PageAttention Exploit in vLLM Inference Engine
CVE-2025-46570

2.6LOW

Key Information:

Vendor

Vllm-project

Status
Vllm
Vendor
CVE Published:
29 May 2025

What is CVE-2025-46570?

The vLLM inference engine is susceptible to a timing attack due to its PageAttention mechanism, which improperly optimizes prompt processing. When a prompt matches a previously processed prefix, the engine accelerates the prefill process, creating significant timing differences. This behavior may be maliciously exploited before the issue was corrected in version 0.9.0. Users are strongly advised to update to the latest version to secure their systems against this vulnerability.

Affected Version(s)

vllm < 0.9.0

References

https://github.com/vllm-project/vllm/security/advisories/...
x_refsource_CONFIRM
https://github.com/vllm-project/vllm/pull/17045
x_refsource_MISC
https://github.com/vllm-project/vllm/commit/77073c77bc200...
x_refsource_MISC

CVSS V3.1

Score:
2.6
Severity:
LOW
Confidentiality:
Low
Integrity:
None
Availability:
Low
Attack Vector:
Network
Attack Complexity:
High
Privileges Required:
Low
User Interaction:
Required
Scope:
Unchanged

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.
CVE-2025-46570 : PageAttention Exploit in vLLM Inference Engine