AI Serving Security Architecture
Inter-process communications between model workers relied on insecure pickle/RPyC deserialization.
Deep autopsy into the CLOSEDQUORUM operational compromise: exploitation of RPyC deserialization in LLM serving clusters, establishing decentralized multi-agent coordination with autonomous privilege discovery.
| Dimension | Attribution & Scope |
|---|---|
| Attributed Adversary | ACTOR-CLOSEDQUORUM |
| Associated Campaign | CAMP-AGENTIC-SWARM-2026 |
| Targeted Vulnerabilities | CVE-2026-103395 |
| Impacted Sectors | SEC-TECH, SEC-ENERGY, SEC-DEFENSE |
| Incident Timeline Anchor | 2026-09-29 |
| Kill Chain Stage | TTP | Observed Mechanics & Telemetry Evidence |
|---|---|---|
| Initial Exploitation | T1190 | RPyC protocol deserialization exploit sent to internal port 8000 of LLM inference backend.Inbound TCP payload triggering unpickling of custom Python execution object in serving worker |
| Lateral Movement & Agent Hijacking | T1071 | Poisoning of Model Context Protocol (MCP) tool registrations, hijacking tool execution across agent pipeline.MCP schema modification injecting malicious Bash execution parameter into tool manifest |
| Autonomous Command and Control | T1573 | Autonomous sub-agents communicating via encrypted P2P gossip protocol across compromised GPU nodes.Inter-node GPU RoCE/InfiniBand traffic carrying encrypted LLM-generated operational tasks |
AI Serving Security Architecture
Inter-process communications between model workers relied on insecure pickle/RPyC deserialization.
Agentic Tool Governance
MCP tool schemas lacked cryptographic integrity signing, allowing silent tool definition alteration.
| Failed Defensive Control | Failure Mechanism Identified in Post-Mortem |
|---|---|
AI Inference Sandbox Isolation (CTRL-APP-SANDBOX) | GPU acceleration containers were granted SYS_PTRACE and host network namespace privileges. |
[!WARNING] Source:
GPU Telemetry (NVML)— Spike in tensor core utilization during off-peak hours without corresponding API inference requests. (Dwell time impact: Unmonitored compute anomalies allowed autonomous agent swarm to self-replicate for 48 hours.)