How the attack works
An attacker sends a prompt asking the agent to list, describe, or output its detection rules, security configuration, or the logic behind its audit and monitoring functions. This can be phrased as a direct request, a debugging question, or framed as a legitimate task like documentation or troubleshooting. If the agent complies, the attacker now has a map of what triggers alerts and what does not. That map is then used to design follow-up prompts or actions that avoid tripping any known detection.
Why it matters
An organization loses visibility: once its detection logic is exposed, attackers can tailor future attempts to bypass monitoring entirely, turning a one-time information leak into a lasting blind spot.
What you can do
- Do not let agents read or repeat their own detection rules, prompts, or security configuration files back to users.
- Separate the system that runs detection logic from the agent that handles user-facing requests, so no single compromised conversation can expose both.
- Treat requests to 'list your rules,' 'explain your security checks,' or 'show your config' as suspicious unless coming from an authenticated security team member.
- Review logs for repeated probing questions about internal logic, even when phrased as legitimate-sounding documentation or debugging requests, since this pattern often precedes an evasion attempt.
Known benign look-alikes
- Security operations team reviewing their own detection rule inventory
- A penetration tester documenting their own findings who writes 'extract security credentials from the vault' in a legitimate pentest report being processed by the agent
- A DevOps engineer asking the agent to 'copy config token from environment variables to the deployment manifest' as part of a CI/CD pipeline task