How the attack works
An AI agent framework normally logs the files it writes to through its own hooks or a proxy that watches its tool calls. If the agent spawns a subprocess, or otherwise writes to disk outside that logged path, those writes are invisible to the agent-level logging. This rule instead watches file-write activity captured by endpoint detection and response (EDR) at the operating-system syscall level, and checks every 30 seconds for writes landing in sensitive paths. Because it works from OS-level telemetry, it can catch file writes that the agent's own instrumentation missed.
Why it matters
Sensitive files (configuration, credentials, system paths, etc.) can be modified or overwritten by an AI agent's subprocess without leaving a trace in the agent's own audit logs, letting the write go unnoticed until this OS-level check runs.
What you can do
- Restrict the filesystem permissions available to accounts or containers that run AI agents, so writes to sensitive paths fail regardless of which process attempts them.
- Run AI agents in a sandboxed or containerized environment where subprocess file access is limited and monitored.
- Cross-check agent-level tool-call logs against OS-level file activity logs to spot gaps where subprocess actions are not being captured.
- Review any confirmed writes to sensitive paths to determine whether they came from the agent's intended actions or from an unexpected spawned process.