Let agents do the work. ActPass holds the authority.
The model can propose a refund, a deploy, an email, or a repo change. ActPass decides whether it's allowed, needs approval, or must be blocked, with signed evidence your auditors can verify. Start with one protected workflow.
Refund agent with bounded autonomy
Risk: An agent reads tickets, touches payment records, and issues refunds: untrusted input, sensitive data, and money movement in one place.
ActPass: Small refunds auto-allow, medium require human approval, over-limit deny. Every decision is bound to a ticket, policy hash, and signed evidence event.
Read the deep dive: Why fintech will buy AI agent security firstCoding agents that cannot push risky changes
Risk: Claude Code and Codex can edit files, run shell commands, push branches, and read secrets.
ActPass: Govern Bash/Edit/Write via PreToolUse hooks; monitor, then enforce; block secret-read + external-comms with the runtime Rule of Two.
Read the deep dive: Govern Claude Code and Codex from inside the chatMCP tool drift caught before it ships
Risk: A tool that was read-only yesterday can quietly gain delete, write, or payment powers tomorrow.
ActPass: Scan, baseline an actpass.lock, and fail the PR when a tool gains high-risk capability. Drift becomes a reviewable code change instead of a runtime surprise.
Read the deep dive: The friendly MCP tool that quietly reads your SSH keysn8n & workflow automation governance
Risk: Workflow nodes mutate tickets, CRM fields, payments, and customer data with no checkpoint.
ActPass: Add a preflight action before risky nodes, require approval for state changes, and record evidence after execution, across MCP, n8n, Slack, Teams, and ServiceNow.
Read the deep dive: How to secure AI agents across MCP, n8n, Slack, Teams, ServiceNow, and compliance toolsWeb agents that read hostile HTML
Risk: An agent that browses the web or reads screenshots ingests attacker-controlled instructions straight into its context.
ActPass: Treat untrusted content as untrusted: classify the exposure, keep at most two of the three dangerous capabilities, and gate outward actions behind policy.
Read the deep dive: How to secure web agents that read hostile HTML and screenshotsRuntime Rule of Two enforcement
Risk: A session that has touched untrusted content AND sensitive data should not be allowed to communicate outward next.
ActPass: Accumulate capability "colors" across a session and block the call that would complete the lethal combination, deterministically and at runtime.
Read the deep dive: Your agent just read a file, hit an API, and got user input. Now what?