Rightful: AI agents ask the owner before a change

A gate for AI help-desk and finance agents. The agent can be fooled, but no password reset or supplier bank-detail change happens until the account owner approves that exact change on their own phone. Measured on Vultr in 5,240 pre-registered runs: without the gate, 0.5% to 44.7% of attack runs changed an account; with it, 0 of 1,199.

Live demoGitHub0

Overview

Attackers talk their way past help desks and finance teams: urgency, authority and a few true details get passwords reset, MFA moved and supplier bank details changed. Clorox claims about $380 million in damages, saying attackers got through its outsourced help desk (the provider disputes it). Business email compromise cost $3.05 billion in 2025 (FBI). Both desks are becoming AI agents, and an agent can be talked into it too. Rightful moves the authority away from the agent. The agent writes Python for each step, and Rightful runs it in a fresh gVisor sandbox with no credentials and no internet. The sandbox's only way out is the gate. When the agent asks for a change, the gate checks a written policy and sends the account owner's own phone a card naming the exact change, with a number to match. Only after approval is that one request forwarded, once. One decline freezes agent changes to that account for 24 hours, so nobody can wear the owner down. The score is not the agent's word: it is the identity server's own log, read by a view-only account. The same gate guards supplier bank details in ERPNext, an open-source ERP. Adding it took a 42-line policy file and no change to the gate. Every conversation can be replayed step by step from the gate's hash-chained ledger, and a signed copy of every record goes to Vultr Object Storage. Try it: one click runs the same attack against the agent alone and against the agent behind Rightful, side by side. No login, no phone needed.

How it uses Vultr

Everything runs on Vultr, and each service has one job: - Vultr Serverless Inference: the agent's model. I tested 3 models on it (qwen3.8-27b, glm-5.3-flash, deepseek-v4.1-flash), called only from the control VM, which holds the key. - Cloud Compute, worker VM (Frankfurt, vc2-2c-4gb): every Python script the agent writes runs in a fresh gVisor sandbox with no credentials, no internet, no cloud metadata and a read-only root. It is destroyed as soon as the script finishes. The worker runs no gate and holds no keys. - Cloud Compute, control VM (Frankfurt, vc2-4c-8gb): the gate, its policy files, Keycloak and the live demo. - VPC + 2 firewalls: the sandbox's only way out is a signed request over the private VPC to the gate. The gate asks the account owner's phone to approve that exact change, and forwards it once. - Object Storage (Amsterdam): a signed copy of every run's record, kept off the demo server. - Vultr API: one command deploys, checks, recycles the worker and tears it all down. The API also created and deleted the evaluation fleets (17 instances created, 15 deleted). Why this split: the machine that runs untrusted, agent-written code holds nothing worth stealing, and the machine that holds authority never runs agent code. Real numbers, all measured on Vultr: 5,240 pre-registered runs. Without the gate, unsafe changes in 0.5% to 44.7% of 1,140 attack runs; with it, 0 of 1,199. For supplier bank details in ERPNext: 24 of 30 without the gate, 0 of 30 with it. Honest trade-off: with the gate, 84% to 90% of honest tickets completed; most of the rest went to a person. Total Vultr bill: $81.84 through 7 Nov.

No comments yet

Be the first to comment on this project.