01 / IDENTITY
Shared passwords
A session on the customer’s credentials can be the customer, an aggregator or an AI agent. To most systems they look the same.
Candor Agent Defense for Fintech / Banks, credit unions and fintechs
An AI agent moved money out of a customer’s account. Which agent was it, and who allowed it?
AI agents are already inside customer accounts, often logged in on the customer’s own password. Candor Agent Defense for Fintech helps institutions see every agent at the account edge, hold each one to what the customer actually approved, stop it when it drifts, and tie every money movement to the agent and the approval behind it.
In build: early access openAgent identity at the account edgeCustomer-approved limitsContainmentAttribution for every money movement
Why now
Banks can see the login. Not who is behind it.
In October 2026, major U.S. banks reported finding unvetted third-party AI agents inside customer accounts. Many were logged in with credentials customers had handed over, and some took actions the customer never approved (reported by Axios on 22 October 2026 and CBS News on 23 October 2026).
02 / AUTHORITY
Beyond what was approved
An agent connected to read balances can try to pay a bill, add a payee or send a transfer.
03 / CROSS-ACCOUNT
Accounts it was never given
A hijacked or careless agent can reach accounts its customer never connected.
04 / MANIPULATION
Hijacked instructions
Injected instructions can turn an approved assistant into a payment nobody asked for.
05 / REVOCATION
Turning one provider off
Cutting off one provider should end every one of its sessions, not just the ones you can find.
06 / DISPUTES
“I did not authorize that”
When a customer disputes a transfer, the institution needs to show which agent acted and under whose approval.
How Candor approaches it
Identify the agent. Check the approval. Contain the drift. Prove every movement.
Candor Agent Defense for Fintech is in build. It brings Candor Agent Defense, whose controls are tested on Candor’s own ranges (results below), to the account edge of financial institutions.
1 / IdentifyWhich agent, from which provider, is in this account?
2 / CheckIs this action inside what the customer approved: accounts, payees, amounts, dates?
3 / ContainAsk the customer, restrict, or stop the agent when it drifts.
4 / ProveA tamper-evident record for every money movement: agent, provider, approval, result.
Verified agent identity
Agents that act in your customers’ accounts carry an identity your institution can check and revoke, instead of hiding behind a customer password.
Customer-approved limits
Each agent is held to what the customer approved: which accounts, which payees, how much and until when.
Unvetted automation found
Automation logged in on a customer’s password is identified, kept away from money movement and recorded.
Inside your environment
Designed to run inside the institution, with customer data kept there. A Candor-hosted option is planned for institutions that prefer it.
A live view for your teams
Security and fraud teams see agents, decisions and money movements as they happen.
Evidence for disputes and examinations
Tamper-evident records that support dispute investigations, board reporting and examination preparation.
Who this is for
Built for the institutions that hold the accounts.
Banks and credit unions
For the CISO and the fraud and risk teams who must answer which agents are inside customer accounts and what they did.
Fintechs and neobanks
For API-first institutions whose customers connect assistants, budgeting apps and payment agents.
Core and digital-banking providers
For the platforms that run online banking for many institutions and can protect all of them at once.
Evidence-first security
The controls underneath are already tested.
Selected results from Candor Agent Defense on Candor’s own test ranges. These are not results from a financial institution; bank-specific results will be published with their dates and environment when they are run.
10 / 10
Delegation-tree revocation
In a compromised-supervisor drill, all ten affected agents lost authority.
18 SEPTEMBER 2026, INTERNAL DRILL
600
Consecutive evaluations
Reported endurance run; 90th-percentile decision latency was 2.4 seconds on the local model.
15 SEPTEMBER 2026, INTERNAL TEST
20 / 20
Revocation and scope scenarios
Revocation, scope and unapproved-destination scenarios contained in a local test range.
19 OCTOBER 2026, LOCAL RANGE
Common questions
AI agents in customer accounts: frequently asked questions
Is Candor Agent Defense for Fintech available now?
It is in build. Banks, credit unions and fintechs can request early access now. Candor confirms scope and dates with each institution directly; nothing on this page is a delivery date.
Does customer data leave the institution?
The design keeps customer data inside the institution. Discovery works from a log slice the institution exports itself, pseudonymized before it leaves: hashed customer ids, no names, no account numbers, no credentials.
Does this block open banking or aggregators?
No. The aim is to tell customer-approved providers apart from unvetted automation and hold each one to what the customer approved, so legitimate providers keep working.
Does it replace our fraud monitoring?
No. It adds an agent layer your current tools do not have: which agent acted, for which provider, under which customer approval. It is designed to work beside your fraud and identity controls.
Does Candor provide a certification?
No certification is claimed on this page. Candor publishes dated internal test results with their environment; coverage is validated in each institution's environment.
Know which agent moved the money, and who allowed it.
Join the early access list for Candor Agent Defense for Fintech.