Skip to main content
FINTECHLAB

FINTECHLAB / AGENTIC FINANCIAL SAFETY

Agentic Financial Safety

FTLAB helps financial institutions evaluate whether AI-agent actions remain faithful to human intent—and design controls that review or block unsafe actions before money moves.

Before an AI agent moves money, institutions need to know whether the action is intended, authorized and safe. FTLAB researches how to make that judgment.

For financial institutions, security teams, regulators and researchers.

Discuss a research partnershipExplore the control-layer modelView the four research tracks

HUMAN INTENT → AGENT BEHAVIOR → FINANCIAL ACTION → LOSS PREVENTION

From human to agent

Financial security now has a second decision-maker.

Financial security used to focus primarily on the human making the decision.

Autonomous agents introduce a second decision-maker.

FTLAB studies both sides of this delegation: the human intent that starts the process and the autonomous behavior that ultimately reaches execution.

HUMAN LAYER

  1. 01

    Human

  2. 02

    Pressure

  3. 03

    Behavioral Change

  4. 04

    Decision

  5. 05

    Loss

AGENTIC LAYER

  1. 01

    Human

  2. 02

    Intent

  3. 03

    Delegation

  4. 04

    AI Agent

  5. 05

    Decision

  6. 06

    Financial Action

  7. 07

    Loss

LIVING MODEL

Before settlement, every autonomous action still has to pass through judgment.

Research tracks

Four directions. One programme.

Agentic Financial Safety is organized around intent, behavior, execution control and machine-speed risk - not a catalogue of agent use-cases.

Status meanings

  • Active — research work is underway with defined questions and methods.
  • Building — infrastructure or tooling is being assembled; not a finished product.
  • Research — exploratory direction under study; findings are not yet settled.
  • Planned — queued for future work; not started.

Does the action still represent what the human wanted?

We study the gap between human intent, delegated authority, agent interpretation and final financial action.

How can we measure whether an autonomous financial action faithfully represents the user's original intent?

  • Intent Fidelity
  • Proof of Intent
  • Intent–Permission Gap
  • Delegated Authority
  • Risk-Based Confirmation
  • Human Approval
  • Intent Integrity

Methodology forthcoming. Detailed methods, experiments and updates will be published as the track matures.

DIAGRAM

  1. Human Intent

  2. Agent Interpretation

  3. Proposed Action

  4. Intent Fidelity

Iₕ = Human Intent · Iₐ = Agent Interpretation · A = Proposed Action · Intent Fidelity = f(Iₕ, Iₐ, A)

Partnership offers

Concrete ways institutions can work with the lab.

Each offer is a research evaluation in a controlled setting — not a product sale and not certification.

Financial-agent red-team research

Teams that need authorized adversarial stress-testing before production rollout.

Request a red-team study →

Live topology

STATUS / INTERACTIVE MODEL

Agents talk. Money moves. Judgment sits in between.

A research sketch of multi-agent financial coordination: message traffic between agents, proposed actions racing toward settlement, and a pre-execution control boundary that can contain, hold, or leave the path open.

Choose a control-boundary mode, then inject a risk packet or click the field. Agents exchange messages on the left; settlement sits on the right; the pre-execution control boundary (membrane metaphor) sits between them.

  • Agents — coordinated actors proposing actions
  • Risk packet — higher-risk financial action
  • Pre-execution control boundary — judgment gate before settlement (visual “membrane”)
  • Settlement — irreversible execution horizon

Control-boundary mode

Simulation actions

Interactive multi-agent topology. Agents exchange messages on the left. A control membrane sits before settlement on the right. Current membrane mode: contained. Click the field or use Inject risk to send a risk packet.

Current state: Contained — Risky actions stop at the membrane. Routine traffic may pass.

Illustrative model — not a live trading system.

Interactive model

STATUS / EXPERIMENTAL

Should this financial action execute?

A conceptual pre-execution control layer. Change the scenario to see how intent, behavior and context can move a decision from ALLOW to REVIEW or BLOCK.

This is an explanatory prototype - not a live risk engine and not production decisioning.

Animated diagram of financial action packets moving from intent through an agent gate. Current gate decision: ALLOW.

SCENARIO

Aligned transfer selected. Gate decision: ALLOW. Behavioral Risk: LOW.

Selected decision: ○ ALLOW

DECISION CHAIN

  1. 01 / HUMAN INTENT

    Pay the listed vendor invoice within the approved limit.

  2. 02 / AGENT BEHAVIOR

    Agent prepares a transfer matching the invoice and mandate.

  3. 03 / PROPOSED ACTION

    Transfer €1,840 to known vendor IBAN.

FTLAB CONTROL LAYER

  • Intent FidelityHIGH
  • Behavioral RiskLOW
  • Context RiskLOW
  • Authority FitHIGH
  • Financial ExposureLOW

GATE DECISION

ALLOWcurrently selectedREVIEWBLOCK

Intent, authority and exposure stay inside the mandate. No behavioral or context anomaly.

Financial intelligence proposes the action. Financial safety decides whether execution should proceed.

Infrastructure

AFSB is the infrastructure used to test the control model above: shared environments, failure scenarios and metrics before real capital moves.

LAB INFRASTRUCTURE · EXPERIMENTAL PLATFORM

AFSB

Agentic Financial Safety Benchmark

A shared experimental environment for measuring financial-agent failures before they become real losses.

STATUS / BUILDING

AFSB is the shared experimental infrastructure for Agentic Financial Safety. It connects the four research tracks through controlled financial environments where agent failures can be observed before real capital is at risk.

No public datasets or scored leaderboard yet. What you see here is the measurement frame under construction.

Next milestone before public release: an agreed run protocol and a first controlled experimental session.

Current stage

Definition and environment design

Completed so far

  • · Research frame across four safety tracks
  • · Initial environment and failure-scenario catalogue
  • · Metric vocabulary for intent, loss and intervention

Next planned release

Shared experimental runs and published protocols — when evidence is ready

Get AFSB updates

Use Contact the lab at the bottom of the page and mention AFSB in your note.

What joining AFSB means

  • Contribute failure scenarios
  • Provide test environments
  • Review metrics
  • Join pilots as they open
  • Receive early protocols
  • Co-author materials when agreed

ENVIRONMENTS

  • Payments
  • Bank Transfers
  • E-commerce

ATTACK / FAILURE SCENARIOS

  • Context Manipulation
  • Prompt Injection
  • Goal Drift

METRICS

  • Intent Fidelity
  • Attack Success Rate
  • Financial Loss

Work with the lab

Study the control layer before execution becomes irreversible.

We partner with institutions that need research on intent fidelity, agent behavior and pre-execution financial controls.

CONTACT THE LABlegal@ftlab.space