tag: Ai-Infrastructure · 6 items
- Engineer — Act: Unpatched critical unauthenticated RCE in LMCache multiprocess mode with no fix available; if running vLLM or any LMCache deployment, immediately disable multiprocess mode or network-isolate the cache server (firewall ZeroMQ port) until a patched release ships.
- SOC/IR — Plan: No IOCs or confirmed active exploitation reported, but organizations running LLM inference infrastructure should create detections for anomalous process spawning or unexpected inbound connections on LMCache/vLLM nodes; start collecting relevant host and network logs from AI compute clusters now.
- Leader — Learn: Highlights emerging attack surface in enterprise AI infrastructure; no immediate leadership action required, but useful context for AI security posture reviews and vendor risk assessments if LLM serving software is in scope.
- Engineer — Plan: If your org runs any externally exposed AI inference or LLM serving endpoints, audit them now for public reachability and missing authentication; review running processes on those hosts for unauthorized miners or scanning activity.
- SOC/IR — Learn: Active campaign to track, but the summary provides no IOCs, ATT&CK mappings, or detection signatures — insufficient detail to write or tune detections today; revisit if a full technical write-up with indicators surfaces.
- Leader — Skip
- Engineer — Act: Public PoC exists for a CVSS 9.8 unauthenticated RCE affecting all Bifrost HTTP transport versions before 2.1.0 — if you run this AI gateway in any environment, patch to 2.1.0 immediately and audit for signs of exploitation in management endpoint logs.
- SOC/IR — Plan: No active exploitation signals yet (EPSS 0.01, not KEV-listed), but a public PoC raises near-term risk — if Bifrost appears in your estate, write a detection for unauthenticated HTTP requests to its management interface before campaigns emerge.
- Leader — Skip
- Signals: CVE-2026-90898 — CISA KEV: not listed, EPSS 0.01, public PoC on GitHub
- Engineer — Plan: Microsoft Threat Intelligence documents active exploitation of exposed LiteLLM gateways leading to credential theft and persistence — no KEV or PoC signal, but if you run LiteLLM or similar AI proxies, audit internet exposure, rotate API keys, and verify no unauthorized processes are running on those hosts.
- SOC/IR — Act: Active attack chain with detectable post-exploitation stages (credential harvesting, persistence, cryptomining) reported by Microsoft TI — pull the blog post for IOCs, then hunt for anomalous processes and outbound connections on any hosts running AI gateway software since the publication date.
- Leader — Plan: AI workloads are now an established attack surface for credential theft and resource abuse; this quarter, ensure AI infrastructure (gateways, API proxies, GPU hosts) is included in your hardening and access-review scope alongside traditional edge assets.
- Engineer — Learn: Academic proposal combining Intel TDX, Intel Trust Authority, and NVIDIA Confidential Computing into a decentralized CVM platform — worth reviewing if you’re evaluating confidential compute options for protecting model weights or training data, but no production tooling or immediate action follows from this paper.
- SOC/IR — Skip
- Leader — Skip
- Engineer — Act: Actively scanning for internet-exposed instances of ComfyUI, Ollama, n8n, Open WebUI, Langflow, and Gradio to harvest AWS keys and Kubernetes tokens — exactly the stack teams deploy fast without firewall controls. Audit now for public exposure of these service ports, restrict to internal networks, and rotate AWS/K8s credentials on any host that ran them exposed.
- SOC/IR — Plan: The TTPs are concrete enough to build detections around: Shodan-driven scanning targeting AI service endpoints, followed by credential exfiltration. Build hunts for unusual outbound traffic or credential API calls originating from AI service hosts; the summary appears truncated so IOCs are not yet available to act on directly.
- Leader — Plan: A claimed harvest of 3,811 AWS keys illustrates the systemic risk of teams rapidly standing up AI infrastructure without security review. Raise with engineering and DevSecOps leadership to establish a deployment standard for AI tooling that includes network isolation requirements before services go live.