NeuralStackSonicWall has released hotfixes for four vulnerabilities in its SMA1000 appliances, which provide remote workers with access to company networks and applications. The most serious, CVE-2026-102255, is a server-side request forgery flaw in the WorkPlace portal that can be reached before authentication and carries a CVSS score of 10.0. SonicWall said an attacker could use the unintended access path to reach internal functionality and perform unauthorized operations, although it reported no evidence that the flaws are being exploited.
The issues affect SMA1000 models 6210, 7210 and 8200v. SonicWall’s advisory did not specify which internal functions could be reached through the SSRF flaw, leaving organizations to treat the pre-authentication exposure as a high-priority issue while applying the available hotfixes.
AI infrastructure faces a separate risk from an unpatched vulnerability in LMCache, open-source software used to accelerate large language model servers such as vLLM. The flaw affects LMCache’s multiprocess mode, where the cache operates as a standalone server and communicates with LLM workers through ZeroMQ. A single network message can execute commands as the user running the LMCache process.
JFrog disclosed the LMCache issue on October 7 and assigned it a severity score of 9.8 when the server is bound to a routable address. The vulnerability, tracked as CVE-2026-105192, affects versions from 0.3.9 through 0.5.5, and no fixed version is available. The risk is greatest when operators configure the service to listen beyond localhost, making it reachable from another machine.
Researchers are also tracking a financially motivated campaign called Canto Incognito that targets exposed AI and large language model infrastructure. The campaign has infected more than 3,400 servers, installing miners including XMRig and Iron and connecting victims to the Kryptex cryptocurrency mining service. Lumen Black Lotus Labs said compromised hosts are reused as scanners and exploit servers to identify and compromise additional vulnerable systems.
The malware used in the campaign is known as PoeLLM because it hides its command-and-control address inside a poem. That technique allows the operators to use compromised AI infrastructure not only for mining but also to expand the botnet, turning exposed systems into tools for further attacks.
Atlassian separately reported exploitation attempts against CVE-2026-21589, a critical arbitrary file access flaw in multiple Data Center products. The vulnerability carries a CVSS score of 9.3 and can allow an unauthenticated attacker to access specific files within the web application root directory, although exploitation requires prior knowledge of the target file’s exact name and path and does not permit directory listing.
The Atlassian flaw affects Bitbucket Data Center, Confluence Data Center, Jira Service Management Data Center, Jira Software Data Center, Bamboo Data Center, Crowd Data Center, Crucible and Fisheye. Exploitation attempts began within two hours of public details becoming available, increasing pressure on organizations that operate the affected self-hosted products to identify exposed instances and apply the relevant updates.
Why this matters
Taken together, the disclosures highlight how attackers can target several layers of modern enterprise infrastructure: remote-access gateways, AI serving systems, and collaboration platforms. The LMCache issue is especially consequential for AI operators because a configuration choice that exposes the service beyond localhost can turn a performance component into a remote code execution entry point. The PoeLLM campaign shows how vulnerable AI systems can also be repurposed to support mining and the discovery of additional victims.
Organizations should prioritize patching SonicWall SMA1000 appliances and affected Atlassian Data Center products, while reviewing whether LMCache multiprocess servers are reachable from routable networks. Until a fixed LMCache version is available, limiting exposure to trusted local interfaces is the clearest risk-reduction measure supported by the available information.