Agent blast radius
Everything an agent could reach if it were compromised.
Agent inventory
A single list of every AI agent, model, key and MCP server you run.
Decision fusion
Combining independent signals into one explainable verdict.
Execution permit
A single-use, scoped token that lets exactly one approved action run.
Human-in-the-loop approval
Holding a high-stakes agent action until a named person approves it.
Indirect prompt injection
Instructions hidden in content an agent reads, not in what the user typed.
Jailbreak
A prompt designed to talk a model out of its safety rules.
Model Context Protocol (MCP)
An open protocol for giving AI agents tools, files and data.
Non-human identity
The credentials and identity an agent acts with.
Pre-action authorization
Checking an agent's tool call against policy before it executes.
Prompt injection
Text that tries to override an AI system's instructions.
Schema drift
When an approved tool's description or inputs change after review.
Shadow AI
AI your organisation runs that nobody registered.
Tool poisoning
Malicious instructions hidden in an MCP tool's description or schema.
+ See it happen
Watch these attacks get decided.
The Simulation Lab runs real attacks through the Intertrace engine, with the reason for every verdict.