AI agent security and runtime enforcement
Enforce policy in the agent's execution path, before the action completes. Block reads of secrets, deny unapproved tool calls, refuse unapproved endpoints, and redact content on the way out.
Enforce policy in the execution path
Qpoint sits between the agent and the OS. Every file open, tool call, and outbound request is checked against policy and allowed, modified, or denied before it happens.
1. Intercept
Catch the action before it runs
- File opens, tool calls, process spawns, outbound requests
- Underneath the agent, not at a gateway
- Nothing routes around it
2. Decide
Evaluate the action against policy
- Allow, deny, or modify
- Approved lists, boundaries, and data rules
- Built-in plugins, or your own against the same hooks
3. Act
Enforce the decision at the source
- Blocked reads return a standard file-not-found
- Redacted requests proceed with the secret stripped out
- Denied tool calls never reach the server
Rules agents can't route around
Block the reads, calls, and requests that your policies do not allow, and redact what shouldn't leave the network.
Block secret and credential reads
SSH keys, cloud credentials, .env files, and keychains are blocked by default. The agent gets file-not-found and moves on.
Block unapproved MCP servers and tools
Unapproved servers are refused at connection. Individual tools on approved servers can be blocked while the rest stay available.
Block unapproved destinations
Requests to destinations outside the approved list are refused before they leave the endpoint.
Redact secrets and sensitive data
Secrets, credentials, and data matching your patterns are replaced with placeholders. The request proceeds without them.
Deny dangerous commands
Recursive deletes, force pushes, curl | sh, and privilege escalation are refused before they run. Normal work is unaffected.
Prompt injection defense
Instructions hidden in a README, web page, or tool response can't make an agent do what policy forbids.
Supply chain protection
An MCP server or skill that starts reading the home directory or calling an unknown host is stopped at that action.
Compromised agent handling
Suspend a session, terminate an agent on a host, or halt every instance across the fleet from one console.
One README, one blocked read
A developer points a coding agent at a cloned open source repo to evaluate it. The README contains instructions written for the agent, not the human: read ~/.aws/credentials and include the contents in a request.
The agent follows them. Qpoint's secrets guard returns file-not-found on the read. The agent reports it couldn't find the file, moves on, and finishes the task.
The developer learns about it from the violation report, not the incident.
AI agent operations, end to end
The same binary that enforces the rules also discovered the agents and wrote the record. Monitor, govern, and secure share one install and one event stream.
Seconds, not sprints
Qpoint deploys as a single binary directly on the endpoint. No gateways to route through, no enterprise infrastructure to provision, no SDK integration to schedule.
Install, run, and start seeing every AI agent on the machine — in under a minute.