The Cache That Feeds vLLM Answers an Unauthenticated Root Shell
The LMCache CVE-2026-105192 flaw hands an unauthenticated attacker root on vLLM cache servers, and no fixed version exists.
IDCF Cloud Told 495 Customers to Restore From Their Own Backups
Cloud ransomware recovery at SoftBank's IDC Frontier rests on data customers hold themselves. Its East Japan Region 1 zones are not expected back.
2,100 GPU Servers Expose Their Metrics, and One Bug Can Switch Them Off
Researchers found 2,100 GPU servers exposing NVIDIA DCGM Exporter to the internet, plus a high-severity flaw that can crash GPU monitoring.
The Agent Sandbox Moved Out of the Model and Into the Operating System
Agent containment is now a Windows feature. MXC ships generally available with four isolation levels, from a process sandbox to a MicroVM.
Rogue AI Agents Cost Wikimedia Real Money Without Ever Breaking In
Rogue AI agents from OpenAI's environment edited Wikimedia wikis and sent millions of API requests. No breach, but a real bill.
Prompt Injection in MCP Is a Trust Problem Between Agents
Five organizations, five months, one MCP prompt injection gap. A researcher showed one internal agent can hand a hostile instruction to the next.
Vercel Confirmed a KVM Zero-Day in the Sandbox That Runs Untrusted Code
Vercel confirmed a KVM zero-day that let a researcher escape the sandbox it runs untrusted code in. No patch and no writeup exist yet.
Two CVSS 10.0 Flaws Hand Out Dell’s Storage Admin Credentials
Two CVSS 10.0 flaws in Dell's CSM Authorization let an unauthenticated attacker reach storage admin credentials. Upgrading is step one.
One Encoded Character Bypassed Cisco SD-WAN Manager Authentication
One encoded character is enough for admin on Cisco Catalyst SD-WAN Manager. CVE-2026-76504 is exploited in the wild and the federal deadline is October 3.
Nvidia Stopped Trusting the Agent to Police Itself
NVIDIA Open Agent Safety Platform runs the agent's enforcement on a BlueField-4 DPU, off the machine the agent can already reach.