Soteria AgentShield: least privilege for autonomous AI
AI agents read email, query databases, call APIs and move money. AgentShield checks every action against policy before it runs.
The agent security problem
An agent's permissions are usually the union of every tool it can reach. One poisoned web page or malicious MCP tool description can redirect the agent to exfiltrate data or take destructive actions, and it will do so with valid credentials.
Core capabilities
Tool-Call Authorization
Identity-aware, context-aware policies decide which tool, which parameters and which data scope are allowed for each call.
MCP Server Security
An allow-list of approved MCP servers, scanning of tool descriptions, and detection of tool poisoning and rug-pull changes.
Human-in-the-Loop
Route high-risk or irreversible actions such as payments, deletions and external emails to an approver through Slack, Teams or email.
Action Tracing
A full trace of reasoning steps, tool calls, inputs and outputs, exported to your SIEM for detection and forensics.
Example policy
Works with your agent stack
- LangChain / LangGraph, LlamaIndex, Semantic Kernel, AutoGen, CrewAI
- Model Context Protocol (MCP) clients and servers
- OpenAI Assistants and function calling, Anthropic tool use, Bedrock Agents
- Custom in-house agent frameworks via SDK