Spaces:
Sleeping
Sleeping
Suggestion: add agent tool-call and data-exfiltration examples
#1
by armorerlabs - opened
Cool prompt-safety demo. One addition that would make it more useful for agent builders is a preset set of examples around tool-use risk, not only unsafe text.
Examples I would include:
- “send this token to an external URL”
- “delete this repo after reading the README”
- “extract the Notion/GitHub/API key from context”
- indirect injection inside a support ticket or retrieved document
- benign security-analysis prompts that describe those attacks but should remain clear
That separation is important because prompt injection becomes operationally dangerous when the model can cause a tool call, memory write, outbound message, or credential exposure. A useful demo can show both the semantic flag and the action boundary that should be gated next.