AI attack field guide

Explore synthetic examples of suspicious AI instructions and benign controls. These scenarios illustrate our local pattern checks; they do not measure a live model’s security.

• A document with a hidden agenda Prompt injection
• A tool result asking for secrets Data exposure
• A dangerous install step Tool misuse
• A request for hidden instructions Prompt injection
• Destruction disguised as cleanup Tool misuse
• Instructions wrapped in encoding Obfuscation
• An unapproved recurring job Tool misuse
• An invisible split in an override Obfuscation
• An ordinary retrieved recipe Benign control
• A normal summarization request Benign control
• Security documentation Benign control
• A regular support article Benign control
• An ordinary calendar request Benign control
• A network status update Benign control
• A support note planting future approval Memory poisoning
• A document planting permanent trust Memory poisoning
• Remembering a harmless display preference Benign control
• A record of one scoped approval Benign control
• A retrieved note planting standing authorization Memory poisoning
• A guide claiming future actions are preauthorized Memory poisoning
• Remembering a handoff format Benign control
• Reviewing standing authorization without storing it Benign control
• Remembering to review vendor permissions Benign control

Privacy & methodology