AI is making identity attacks cheaper and more convincing: better phishing to steal passwords, cloned voices and faces to impersonate colleagues, and automated social engineering aimed at help desks and approval processes. The defenses that hold up are phishing-resistant sign-in (passkeys and hardware security keys), MFA everywhere, tighter account recovery, and a rule that sensitive requests are verified through a second channel no matter how real the voice or face seems. AI identity security is less about new tools and more about removing the weak points AI is best at exploiting.

This post explains the main ways attacks on accounts have changed, why some common forms of MFA are no longer enough on their own, what passkeys are in plain English, how to protect the help desk and approval processes, and a practical upgrade path for a small business.

How are attackers using AI against logins and accounts?

The goal of most attacks on businesses is still simple: get into an account that matters, such as email, banking, payroll, cloud storage or an admin console. AI helps attackers at almost every step.

AttackHow AI helps the attackerDefense that holds up
Credential phishingFluent, personalized emails and fake login pages in any language, at scalePasskeys or security keys, which won't work on a fake site
Real-time phishing proxiesAutomated kits that relay your login and one-time code to the real site as you typePhishing-resistant MFA; sign-in rules based on device and location
MFA fatigueRepeated push prompts, plus a convincing "IT support" message telling you to approveNumber matching or passkeys; alerts on repeated prompts
Help desk social engineeringCloned voices and researched backstories to request a password or MFA resetStrict identity verification before any reset
Executive impersonationVoice or video deepfakes asking staff to approve payments or share accessOut-of-band verification and two-person approval
Password guessingGuesses tailored from personal details posted onlineUnique random passwords from a password manager; breached-password checks
Biometric spoofingSynthetic faces or voices aimed at selfie checks or voice authenticationLiveness detection, multiple factors, no voice-only authentication for sensitive actions

The common thread: anything that relies on a person recognizing something as real, such as a familiar voice, a believable email or a convincing login page, is weaker than it used to be. Controls that work regardless of how convincing the attacker is are the ones to invest in.

Why isn't ordinary MFA enough anymore?

Multi-factor authentication (MFA) means proving who you are with more than a password, usually a code or a prompt on your phone. It remains one of the most effective security measures a business can take, and any MFA is far better than none. But not all MFA is equal.

  • SMS codes can be intercepted or redirected through SIM swapping, where an attacker convinces a mobile carrier to move your number to their SIM.
  • One-time codes from an app are stronger, but a phishing page can ask for the code and pass it to the real site within seconds. Automated kits that do this have become common.
  • Push approvals ("Is this you? Approve / Deny") can be abused by sending prompts repeatedly until someone taps approve to make them stop, especially if a fake support message tells them to.

Phishing-resistant methods solve the core problem: they cryptographically check that you're on the real website, so a fake login page gets nothing usable. That's what passkeys and hardware security keys do.

What are passkeys, in plain English?

A passkey replaces a password with a pair of digital keys. One key stays securely on your device (phone, laptop or a physical security key); the other is held by the website. When you sign in, your device proves it holds the matching key, usually after you unlock it with your fingerprint, face or PIN.

Why that matters for businesses:

  • Nothing to steal in a phishing email. There's no password or code for a user to type into a fake page.
  • Bound to the real site. A passkey for your real email provider simply won't respond to a lookalike domain.
  • Easier for staff. Unlocking a phone is quicker than typing a password and a code.

Support for passkeys has grown across major email, productivity, identity and banking platforms, though coverage varies by service and plan, so check what your own tools support. Hardware security keys offer the same protection in a physical form and are a good fit for admins and finance staff.

How do you protect the help desk and account recovery?

Attackers increasingly skip the login page and go to the people who can reset it. With a cloned voice and details scraped from social media, a caller claiming to be a locked-out employee can be very persuasive.

Tighten recovery with rules like these:

  1. No resets based on voice or caller ID alone. Voices can be cloned and caller ID can be faked.
  2. Verify through a known channel. Call back on the number in your HR system, or confirm through the employee's manager on a separate channel.
  3. Use in-person or video verification with a known face for high-privilege accounts, and remember that video can be faked too, so pair it with something else.
  4. Notify the account owner whenever MFA or a password is reset, through a channel the attacker doesn't control.
  5. Extra scrutiny for admin and finance accounts, including a waiting period or second approver for resets.

If you're automating parts of your IT help desk with AI, keep identity-sensitive actions such as resets and access grants behind strict verification and human approval. Our post on AI IT helpdesk automation covers where that line sits.

How should you verify sensitive requests now that voices and faces can be faked?

Treat any request to move money, change bank details, share credentials, grant access or send sensitive data as unverified until confirmed through a separate channel you initiate.

  • Call back on a number you already have, not one given in the message.
  • Use a shared code word or verification question for urgent executive requests, agreed in person and not stored in email.
  • Require two people for payments to new accounts and changes to supplier bank details.
  • Make delay acceptable. Leadership should state openly that they expect to be verified and won't be annoyed by it.

Our post on deepfakes and AI voice scams goes deeper on impersonation, and our guide to spotting AI-written phishing emails covers the email side.

What about the identities of your AI agents?

As businesses deploy AI agents that log into systems, send emails and update records, those agents have identities too: service accounts, API keys and access tokens. These "non-human identities" are increasingly attractive targets, because they often have broad access and nobody watches them closely.

Treat them like privileged employees:

  • Give each agent its own account, never a shared human login.
  • Grant the minimum permissions it needs, and prefer read-only where possible.
  • Store keys in a secrets manager, not in code, spreadsheets or prompts.
  • Rotate credentials and revoke them immediately when an agent is retired.
  • Log every action the agent takes under its identity.

This is part of the wider set of AI agent guardrails we put around any agent with access to business systems.

What's a practical upgrade path for a small business?

You don't need to change everything at once. In rough order:

  1. MFA on every account that matters, starting with email, finance, payroll, cloud storage and admin consoles.
  2. A business password manager for everyone, with unique passwords and no shared logins in spreadsheets.
  3. Move off SMS codes to an authenticator app with number matching, as a stepping stone.
  4. Passkeys or security keys for admins, finance and executives, then for everyone where your tools support it.
  5. Written recovery and verification rules for the help desk and for payment and access requests.
  6. Sign-in policies that block or challenge risky logins, using the risk detection built into many identity platforms.
  7. An inventory of agent and service accounts, with owners and minimal permissions.

For the wider set of tools, see our guide to which AI security tools a small business actually needs.

Frequently Asked Questions

Can AI crack passwords?

AI can make guessing smarter, for example by generating likely passwords from personal details posted online, but long, random, unique passwords from a password manager remain impractical to guess. The bigger risk is passwords being phished or reused across sites, which passkeys and MFA address.

Are passkeys safe for a small business to adopt?

Yes, and they're often easier for staff than passwords plus codes. Plan for recovery: make sure each person has a backup method or a second passkey, and decide who can help when someone loses a device. Check which of your tools support passkeys before rolling them out.

Is voice authentication still safe?

Voice alone shouldn't be trusted for sensitive actions, because convincing voice clones can be made from short recordings. If a service uses voice recognition, treat it as one factor among several, and never let a voice on the phone be enough to reset an account or approve a payment.

Should we use SMS codes or nothing?

SMS codes are far better than nothing. If they're all a service offers, use them. Where an authenticator app, passkey or security key is available, switch to that, prioritizing email, finance and admin accounts.

How do we protect AI agents' credentials?

Give each agent its own account with minimal permissions, store its keys in a secrets manager rather than in code or prompts, log its actions, and revoke access when it's retired. Review agent accounts regularly as you would any privileged user.

Want help tightening logins, approvals and agent access?

We help businesses build approval workflows, verification steps and properly scoped integrations for both people and AI agents through our enterprise software work. Tell us how sign-ins and approvals work today and we'll suggest the upgrades that matter most for your risk.