Race in vLLM's prefill/decode split lets a stale KV-cache read leak data and crash the service
The flaw is in KV-block lease management between the prefill and decode components. When a request stays in the orchestrator longer than the default lease, the producer reclaims and reuses the KV-cache block while the consumer still reads with an expired remote block ID. That race makes the consumer silently load another request's hybrid attention state or recurrent state. An attacker can mix up data this way, or craft input that triggers a CUDA illegal-address error and crash the whole inference engine, seriously damaging confidentiality, integrity, and availability.
Component
vLLM is a high-performance inference framework for large language models. It optimizes Transformer attention computation and memory management, and uses a distributed architecture that separates prefill from decode to maximize GPU utilization. Through low-level communication libraries such as NIXL it transfers tensors and synchronizes state across devices, supporting high-concurrency, low-latency enterprise AI applications.
Add a strict validity check for prefill-block leases at the data-transfer layer so the consumer verifies that the lease has not expired before reading. Add heartbeat lease renewal so long-waiting requests stay valid, and fail the operation when an invalid read is detected instead of continuing, eliminating the race.
CSSA Exclusive Early Warning
10.0 Critical
An authorization flaw in the litellm proxy lets a remote attacker obtain unrestricted access and damage system integrity
granted_scope does not distinguish a missing value from an empty string when parsing the scope claim in an ID-JAG assertion, so an empty scope is incorrectly mapped to an unrestricted global credential. An attacker can send a network request with a validly signed token that contains an empty scope, bypass route-based fine-grained access control, and call every LLM and MCP inference interface behind the proxy without restriction, severely damaging confidentiality and integrity.
Component
litellm is a unified proxy for large-language-model calls. It hides differences among underlying models and provides a standard API, covering routing, key management, and request authentication. Identity assertions such as ID-JAG control access to microservices or MCP resources and isolate privileges.
Strictly distinguish whether a claim field is present from whether its contents are valid. If scope is not None but empty or whitespace, return an invalid-scope rejection. Grant the default credential only when the assertion omits the field entirely, and add regression tests for the empty-string case so the authorization logic stays robust.
CVE Intelligence
2
CVE-2026-113363
CVSS
10.0 Critical
2026-11-09
Remote command execution in NEC WebSAM WinShare runs arbitrary code as SYSTEM
In NEC Corporation WebSAM WinShare 7.2.0 through 7.5.0, 7.6.0, and 7.7.0 through 7.7.2, the network-communication module on Windows has improper input validation (CWE-749). An attacker can send a crafted network packet to a computer running WinShare and trigger an unauthenticated remote attack that executes arbitrary commands as SYSTEM. Every WinShare user who has not updated to a fixed version is affected, including enterprise remote-access deployments. The attacker needs no authentication, can attack remotely, and directly obtains the highest system privileges.
Component
NEC WebSAM WinShare is a remote-access and session-management solution for Windows that lets users reach internal resources securely over the network.
Risks
Complete system control: Commands executed as SYSTEM give full control of the compromised host, including installing malware, stealing sensitive data, or damaging system integrity
Unauthenticated remote attack: No credentials or user interaction are required; a crafted network packet is enough, which greatly lowers the barrier
Lateral-movement risk: With SYSTEM privileges the attacker can move laterally inside the network and threaten the whole enterprise network
Deploy an intrusion detection system (IDS) at the network boundary to monitor WinShare ports for abnormal traffic and crafted packets
Restrict network access to the WinShare service so only trusted IP addresses can connect
CVE-2026-124450
CVSS
9.9 Critical
2026-11-08
Path traversal in the OpenEMR fax module leads to remote code execution
Before OpenEMR 8.3.0, the disposeDocument action in the fax/SMS module fails access control and does not validate the file_path parameter, a path-traversal flaw (CWE-73). An authenticated low-privilege user can read sqlconf.php to recover database credentials and write arbitrary files as the web user, leading to remote code execution. Every OpenEMR user who has not updated to 8.3.0 is affected. A low-privilege account is enough to trigger the flaw and fully control the server with a crafted request.
Component
OpenEMR is open-source electronic medical record and medical practice management software, widely used by healthcare organizations for patient data and clinical workflows.
Risks
Database credential disclosure: The attacker can read the configuration file for database connection details and further penetrate the backend database
Arbitrary file write: The attacker can write files anywhere on the server as the web user, laying groundwork for further attacks
Remote code execution: By writing a malicious script or configuration file, the attacker can run arbitrary system commands and fully take over the server
Low-privilege exploitation: An authenticated low-privilege account is enough to launch the attack, lowering the barrier