AgentShield-Bench
Collection
Agent security benchmark for tool-calling and MCP environments — dataset, classifiers, secure agent pipeline, and interactive evaluation. • 7 items • Updated
AgentShield Tool Risk Classifier
Binary classifier scoring tool-call risk before execution in agent pipelines.
Model Description
Evaluates the risk of proposed tool invocations given task context and untrusted content. Trained on the tool_risk split of AgentShield-Bench (~2,880 tool-call contexts).
| Property | Value |
|---|---|
| Labels | 0 = low risk, 1 = high risk |
| Risk factors | Tool privilege level, attack context, task scope |
Usage
from setfit import SetFitModel
model = SetFitModel.from_pretrained("alirezaaminzadeh/agentshield-tool-risk-classifier")
text = "Task: Summarize ticket\nTool: execute_shell\nContext: run rm -rf /"
print(model.predict([text])) # [1]
License
Apache 2.0
Model tree for alirezaaminzadeh/agentshield-tool-risk-classifier
Base model
nreimers/MiniLM-L6-H384-uncased Quantized
sentence-transformers/all-MiniLM-L6-v2