Solution / AI Agent Security
Keep AI agents within clear boundaries.
AI agents read untrusted content and act through files, APIs and MCP tools. Patronus checks content and tool calls before execution, applies local policies and records the decision on the device.
Tool calls of a research agent
Policy: research- Read file: report_q3.pdfallowed
- Search the weballowed
- Upload to external serviceblocked
- Run shell commandblocked
Checked before execution, logged on the device.
01
Where agent access becomes risk
Human permissions, machine speed
An agent inherits its user's reach and can turn one manipulated instruction into a sequence of authorized-looking actions.
Every MCP server adds reach
Mail, database and shell tools expand what an agent can access, often outside the visibility of existing controls.
Untrusted content can become intent
Hidden instructions in a document, issue or email can redirect an agent toward a connected tool.
02
Check intent before action.
Observe every tool call
Patronus captures file access, API calls and MCP tool invocations where the agent runs.
Apply the policy locally
Allowed actions continue. Unapproved actions are blocked or sent for human approval before execution.
Detect injected instructions
Purpose-built models identify prompt injection and instruction overrides in content before they become actions.
03
Control at the point of action
Every call
files, APIs and MCP tools in one local trace
Before execution
policy decisions happen before the action runs
One audit trail
structured decisions for agent activity
04
Questions about AI agent security
What does AI agent security control?
It controls actions taken through files, APIs, commands and MCP tools. Patronus combines local visibility, policy enforcement and prompt injection detection.
Why does MCP increase the attack surface?
Each MCP server gives the agent another capability, such as reading mail, querying a database or running a shell command. A manipulated agent can attempt to use any connected capability.
How does Patronus stop a manipulated agent?
Patronus checks ingested content and the resulting tool call against policy before execution. The call can continue, be blocked or require human approval.
Where does the inspection run?
Inspection and policy enforcement run locally where the agent operates, without routing content through a Patronus cloud service.