[KongchangAI]
Product

thoughtpay

A semantic verification tool for AI agent payment security. It reads an agent's reasoning trace and checks its consistency with the task to intercept payment requests manipulated by prompt injection, issuing an APPROVE or BLOCK verdict before the payment reaches the payment gateway.

Timeline (last 90 days)

Oct 12

thoughtpay reads the Agent's reasoning trace to determine whether a payment can be logically derived from the task the Agent was set, thereby blocking the payment.

Unverified50%
Oct 12

thoughtpay's workflow includes: the Agent reasons normally and collects the reasoning trace step by step; when the Agent calls the pay() tool it is intercepted; the classifier gives an APPROVE or BLOCK verdict based on the task description, payment policy, reasoning trace and payment request, along with a consistency score, risk signals and an audit ID.

Unverified50%
Oct 12

thoughtpay's classifier is played by Claude.

Unverified50%
Oct 12

When injected instructions create a logical break in the Agent's reasoning, the consistency score drops sharply and the payment is intercepted before reaching Stripe.

Unverified50%
Oct 12

The thoughtpay integration relies on two components: a callback handler responsible for collecting reasoning text while the Agent runs, and a guarded pay tool that hands the reasoning trace to thoughtpay for verification before execution.

Unverified50%
Oct 12

PaymentPolicy configuration includes agent_id, task_description, allowed_vendors, allowed_categories, max_amount_usd, allow_wire_transfers and require_known_counterparty.

Unverified50%

All Facts (6)

Source Articles