Six Habits to Stop AI-Powered Fraud and Data Leaks

Crew bank details, tax forms, and payment approvals pass through more accounts and vendors than ever, and attackers now have AI to help them reach that information. Cloned voices, synthetic video, and polished, personalized messages make fraudulent requests harder to tell from real ones.

October’s global security incidents show how fast that shift is moving. AI helped researchers uncover critical flaws in PHP, and experts warn that attackers can use the same capabilities to find and exploit weaknesses faster than teams can fix them. The weak spots that reach production finance are familiar ones: stolen credentials, trusted vendors, and people under time pressure. AI can make each one easier to exploit. This post covers what happened in the industry and six habits that help protect your payments, crew data, and accounts.

What happened

The incidents came in a tight cluster. On October 21, PHP released emergency patches for two critical flaws found with AI-assisted analysis, and some government websites in several European countries went offline temporarily during updates. Reporting so far shows no single coordinated attack behind the outages. The same day, attackers used stolen GitHub credentials to push a malicious version of LiteLLM, an open-source gateway that connects applications to AI models, carrying credential-stealing malware. On October 27, the AI development environment Warp disclosed that a third-party vendor breach exposed agent logs and telemetry.

AI agents carry their own risk. Earlier this year, researchers showed that hidden instructions in an ordinary web form could trick agents in Microsoft Copilot Studio and Salesforce Agentforce into leaking data. Both vendors patched the flaws, and Salesforce pointed to human approval of sensitive actions as one mitigation.

Attackers are also using AI against people, pressuring employees into approving transfers, resetting credentials, or sharing customer data.

Six habits that protect your production

The following habits address the weak points the incidents above exposed, and most take little time to put in place.

1. Verify payment and bank detail changes outside the original channel

Confirm any change to payee bank details, and any urgent payment request, through a phone number you already have on file or an in-app workflow. Never use the channel the request arrived on. A familiar voice or face on a call no longer counts as verification, so treat urgency, secrecy, and "skip the usual approval" as signals to slow down. Require a second approver for large or first-time payments. In Wrapbook, approval flows let finance teams build that second review into the payment process.

2. Strengthen sign-in security

Start with the accounts that move money or hold crew data: administrator, finance, payroll, and email accounts. Use passkeys or hardware security keys where they're available, as verification codes sent by text or email can be intercepted or phished. Use a password manager, and never share a verification code or approve a sign-in you didn't request.

3. Limit what AI tools can see and do

Before your staff starts to use a new AI tool, confirm that the tool is approved by your company and check how it stores and uses data. The same scrutiny applies to AI browser extensions and meeting assistants. Don't give an AI agent write access to payroll, banking, or email unless a person approves each sensitive action. Treat anything an agent reads, including emails, forms, web pages, and uploaded documents, as untrusted, since hidden instructions inside that content can steer the agent. Always validate outputs you receive from AI.

4. Protect keys and secrets

This one applies if anyone on your team, or a vendor you work with, builds with AI tooling. Keep API keys out of code, prompts, and chat. Store them in a secrets manager, use a separate key for each tool so one leak has limited reach, and pin dependency versions so updates are reviewed before they're installed. After any vendor breach notice, rotate the keys and tokens involved. For example, anyone who installed LiteLLM version 1.107.0 should assume their credentials were exposed and rotate them.

5. Review who and what has access

Give employees, freelancers, vendors, and connected applications only the access they need. Productions wrap and crews turn over constantly, so remove access as soon as an engagement ends. Every few months, review the integrations connected to email, file storage, and financial systems, and ask your vendors which AI tools handle your data. When a vendor sends a breach or security notice, verify it through their official website or support channel before following any instructions in it.

6. Make reporting easy and have a plan

Give your team one clear contact for suspicious messages, unexpected login prompts, accidental disclosures, and questionable payments, even if they've already clicked something. Write a one-page response plan that names who to call: your bank, Wrapbook support, and the FBI's Internet Crime Complaint Center (IC3). Speed matters most when a payment has been redirected, so report suspected fraud immediately.

Wrapping up

AI is changing how quickly attacks can move. The habits that hold up under pressure are the practical ones: verify payment changes, protect sign-ins, control what AI tools can see and do, and report quickly.

If you suspect fraudulent activity on your Wrapbook account, contact us at support@wrapbook.com immediately.