NVIDIA AI Red Team Publishes Mandatory Security Controls for Agent Sandboxing — Configuration Files Are the Kill Chain

NVIDIA Technical Blog·high signal

NVIDIA's AI Red Team released practical guidance declaring that LLM-generated code must be treated as untrusted output. Mandatory controls: block network egress, block file writes outside workspace, block ALL writes to configuration files (hooks, MCP definitions, skill scripts execute outside sandbox contexts). The key insight: indirect prompt injection via malicious repos is the primary threat vector, and config file modification is the escape hatch.