REPRODUCIBLE AI SECURITY
ToolShield Results Explorer
A static view of the committed prompt-injection benchmark. No model key, backend, tracking, or submitted prompts.
GENERALIZATION PROFILE
How models compare
ROC-AUC for the selected protocol. The tool-holdout view exposes the harder unseen-tool case.
DETECTOR → ENFORCEMENT
Risk-aware decision simulator
The same detector score can be acceptable for a read operation and blocked for a privileged action.
REPRODUCIBILITY PROVENANCE
Evidence fingerprints
What the protocols test
- S_random
- Familiar attack and tool patterns.
- S_attack_holdout
- Generalization to an unseen attack family.
- S_tool_holdout
- Generalization to an unseen tool.