highexperimentalLinuxOCIT1565.001

Linux Agent Overwriting Prompt Template Or System Instruction Files

Detects AI agent processes modifying prompt template files or system instruction configurations, which represents a direct attempt to alter the agent's core behavioral guidelines.

Updated Jan 10, 2025 · OCI AI Security Team

agentic-aimemory-poisoningprompt-templatesystem-promptlinuxociowasp-asi06

Problem Statement

System prompt and instruction files define the safety boundaries and behavioral guidelines for an AI agent. Unauthorized modification of these files removes safety guardrails and can completely redirect the agent's behavior, making it a persistent insider threat within the infrastructure.

Sample Logs

{"timestamp":"2025-01-10T11:00:00Z","computer_name":"oci-worker-25","user":"agent_svc","image":"/usr/bin/python3","target_filename":"/opt/agent/config/system_prompt.txt","event_type":"ModifyFile"}

Required Fields

image

target_filename

user

computer_name

False Positives

·Legitimate agent configuration management systems that update system prompts as part of a controlled deployment
·A/B testing frameworks that rotate prompt templates

Tuning Guidance

Make prompt template files immutable in production using chattr +i or filesystem read-only mounts. Alert on any modification regardless of source in production environments.