Can an AI Agent Refund $1,200 Without Permission? - Aaron Tainter | init() 2026 @WorkOS
Can an AI Agent Refund $1,200 Without Permission? - Aaron Tainter | init() 2026  @WorkOS
Uploaded October 2026 | Updated November 2026, 4 days ago
Airlock is an authorization system between an organization's agents and its data integrations. For developers building agent workflows, the demo follows a billing investigation in Cursor across Linear, Snowflake, Stripe, and Gmail, with integrations running on WorkOS Pipes.

Aaron shows how intent tokens and centralized policies govern integration calls beyond broad API scopes. He walks through activity monitoring, token revocation, deterministic enforcement rules, and runtime rules evaluated by an LLM-based governing agent.

The billing workflow demonstrates three policy outcomes: a $1,200 refund is denied, a $300 refund requires human approval, and an email containing financial data is blocked.

Chapters:

See how Airlock handles agent authorization for your apps and services by managing API keys and user tokens securely.

Managing access for autonomous agents requires strict controls. This demo shows how to implement authorization policies that govern what tools and services your agents can access. We walk through a practical scenario involving an invoice discrepancy for Acme, demonstrating how to trace activity across Stripe and Snowflake.

You will see how the system enforces API key management to ensure that sensitive data integrations remain secure. The workflow highlights the importance of human approval when agents take actions like processing refunds, ensuring that automated tasks align with organizational standards.

Subscribe for more technical breakdowns on agent security, and let us know what infrastructure challenges you are solving in the comments.

workos.com/airlock

0:00 Introducing Airlock
1:00 Billing investigation in Cursor
2:07 Activity and intent tokens
3:29 Testing refund limits
5:00 Configuring Airlock rules
6:04 Deterministic and runtime policies
7:17 Human approval workflow
8:36 Blocking financial data emails
9:44 API and MCP support
Can an AI Agent Refund $1,200 Without Permission? - Aaron Tainter | init() 2026How AI Agents Are Replacing Configuration — David Crawshaw, exe.dev | Agent NightEnterprise Authorization Patterns for AI Agents — Garrett Galow, WorkOS | MCP NightCan Your AI Survive Crashes? Fault-Tolerant Gen Code — Maxim Fateev, Temporal | HumanX 2026AI Is Rewriting Healthcares Hidden Layer — Zack Kanter, Stedi | re:Invent 2025AI Will Change Jobs Faster Than You Think — Aaron Levie (Box) & Michael Grinich | init() 2026Why Autonomous Agents Still SuckHow I Gave My ESP32 Gadget a Secure Login - chantastic | init() 2026Agent Night: Is AGI Already Here? Jaya Gupta, Flo Crivello & swyx on AI AgentsLower Your Incident Cost to Ship Faster — Chris Evans, Incident.io | re:Invent 202510x the Tokens, Only 2x the Output — Nicholas Arcolano, Jellyfish | AI Engineer Worlds Fair 2026Dwarkesh Patel | Dwarkesh Unplugged presented by WorkOS
WorkOS |

Can an AI Agent Refund $1,200 Without Permission? - Aaron Tainter | init() 2026

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER