Rohan Raghu / Countersign All work Contact
Startup · in development

Countersign

A second signature on every move an AI agent makes with money.

AI agents are starting to read invoices, answer customer email and act on financial news, then pay, trade and move data on their own. Every document they read is a chance for someone to slip in an instruction. Countersign checks what an agent reads, and what it does next, before the money leaves.

StatusBuilding
Next milestoneIndependent evaluation
StartedOctober 2026
RoleFounder
StackPython · PyTorch · ONNX · FastAPI
One session, three decisions: a hidden instruction in an invoice is flagged, the bank-detail change it asks for is held, and the payment to the changed account is blocked.
What it is

Two checks, one API.

01

Scanner

POST /v1/scan

Reads untrusted content before the agent does and looks for instructions aimed at the agent: hidden in HTML, zero-width characters, look-alike letters, footers or another language. It returns a risk score and points at the exact span it flagged.

02

Monitor

POST /v1/events

Watches the agent's sequence of actions and answers allow, hold or block. A single message can look harmless; a payee change right after reading an outside document, followed by a payment to that payee, does not.

Why

General guardrails ask whether one piece of text looks malicious. Fraud in finance usually looks like ordinary paperwork, one step at a time.

Countersign is built around a finance-specific threat model: accounts-payable agents that pay invoices, fintech copilots that read customer email, and research agents at funds that read filings and news. Besides "is this text an attack?", it asks "should this agent be allowed to do this, given everything it has read in this session?"

It also comes out of the same lesson as my honest backtester: a number that looks good on the data you tuned it on means very little. Security tools are easy to make look perfect in a demo, so the evaluation here is designed to be hard to fool, starting with the model's own author.

Progress

Where the build is.

Milestones in the order they are being built. This page is updated as each one lands.

0 of 0 milestones done
Done In progress Up next

The milestone list needs JavaScript to display.

    Build log

      Last updated –

      How it is built

      Rules I am holding the project to.

      01

      Judge it on attacks it has never seen

      Thresholds are set on a validation split only. Results are reported separately for held-out attack types, document types and hiding techniques, because in-distribution scores flatter every model.

      02

      Fast enough to sit in the loop

      A check that slows the agent down gets turned off. The scanner is built to run on an ordinary CPU in milliseconds, not on a GPU behind a queue.

      03

      Private by default

      Countersign stores hashes and lengths of what it scans, not the raw text, unless a customer explicitly turns that on.