ForgeAI / TOFU / yubikey for business ai access

What Is a YubiKey, and Why Would a Business Use One for AI Access?

What a YubiKey does, how FIDO2 hardware keys stop phishing, and why businesses use one to secure access to AI systems holding sensitive data.

Published
Reading time
6 min read
  • YubiKey
  • phishing-resistant authentication
  • AI security
  • ForgeAI
  • access control

A YubiKey is a small physical device, usually shaped like a USB stick, that proves who you are through public-key cryptography instead of a password or a text-message code. You plug it in or tap it against your phone, and it signs a challenge from the service you're logging into using a private key that never leaves the device. For a business running AI tools on sensitive data, that distinction matters: a stolen password can be typed into a fake login page, but a cryptographic signature tied to a physical key generally can't be phished the same way.

What Is a YubiKey, Exactly?

A YubiKey is a hardware authenticator made by Yubico that implements open standards from the FIDO Alliance, specifically FIDO2 and its browser component, WebAuthn (Web Authentication API). When you register a YubiKey with an account, the device generates a public-private key pair. The public key goes to the service; the private key stays locked inside the YubiKey's secure hardware and never touches the network, the browser, or even the device it's plugged into.

Logging in works like this: the service sends a random challenge, the YubiKey signs it with the private key after you touch the metal contact (a physical presence check), and the service verifies the signature with the public key it already has on file. No password travels across the connection. No one-time code gets typed into a browser window that could belong to an attacker.

How Does a YubiKey Protect Against Phishing?

Passwords and SMS or app-based one-time codes share a weakness: a user can be tricked into handing them over. A convincing fake login page captures the password or the six-digit code, and the attacker replays it on the real site within the code's short validity window. This is how most credential-phishing and "adversary-in-the-middle" attacks succeed even against accounts with traditional multi-factor authentication (MFA) enabled.

FIDO2/WebAuthn closes that gap through domain binding. The cryptographic signature a YubiKey produces is tied to the exact origin (the domain) that requested it, so a credential registered for your company's real login page won't work on a look-alike phishing domain, even a nearly identical one. The U.S. government's own guidance on this is direct: the General Services Administration's phishing-resistant authenticator playbook states that manual entry of a code, meaning an OTP typed by hand, inherently fails to meet the phishing-resistance bar, while FIDO authenticators and PIV smart cards do meet it (idmanagement.gov (opens in a new tab)).

NIST, the federal government's standards body, codifies this in Special Publication 800-63B, its digital identity guidelines. At Authenticator Assurance Level 3 (AAL3), the highest tier NIST defines, an authenticator "SHALL have a non-exportable private key and SHALL provide phishing resistance," with keys held in an isolated, hardware-protected execution environment (NIST SP 800-63B (opens in a new tab)). A YubiKey is one of the small number of devices that meets that bar out of the box.

Why Would a Business Use a YubiKey for AI Access Specifically?

AI tools raise the stakes on access control in a way ordinary software often doesn't, because the whole point of connecting a large language model to your business is to let it read and reason over what you know: contracts, financial records, patient notes, source code, internal strategy documents. A compromised login to a general SaaS tool exposes whatever that one tool holds. A compromised login to an AI system that's been given broad read access to your file server or your knowledge base can expose all of it in one conversation.

The OWASP Foundation's Top 10 for LLM Applications names this directly. "Excessive Agency" (LLM06) covers systems granted more operational autonomy or data reach than a given task requires, and "Sensitive Information Disclosure" (LLM02) covers exactly the scenario above: an AI system trained on or connected to confidential data leaking it to whoever can get in (OWASP GenAI Security Project (opens in a new tab)). Neither risk is caused by the AI model itself being unreliable. Both are caused by weak access control around a system that was, by design, given a lot to look at.

That's the practical argument for hardware-key authentication on AI systems: if the account that gets phished can only reach one narrow tool, the blast radius is small. If it can reach a system indexed across your whole operation, it isn't.

How Does ForgeAI Use a YubiKey for Access Control?

ForgeAI, Monarch Cyber Solutions' on-premise AI platform, builds a physical YubiKey requirement directly into its access model rather than treating hardware authentication as an optional add-on. According to ForgeAI's own product documentation, the system requires "a physical YubiKey token as part of the access model," and the stated principle is blunt: no authorized key, no system access (ForgeAI: Why ForgeAI (opens in a new tab)).

That fits the broader design goal behind running AI on-premise in the first place: keeping documents, queries, and model responses inside a network the business controls, rather than routing them through a third-party cloud service. Hardware-key access control is the login-side half of that same idea. Data locality limits where information can travel once it's inside the system; phishing-resistant authentication limits who can get into the system to begin with.

ForgeAI's documentation is also careful to note that a physical key alone isn't a complete security program. Organizations still need to define who's authorized, what approval process changes access, and how output gets reviewed once people are logged in. A YubiKey answers one question well: is this really the authorized person logging in. It doesn't answer questions about what that person should be allowed to do once they're inside.

Is a YubiKey the Same Thing as Two-Factor Authentication?

Not quite, though it's often grouped under that umbrella. Traditional two-factor authentication (2FA) usually pairs a password with something else: a text message code, an app-generated code, or a push notification. All three of those "something else" factors can be phished, because they're either shared secrets a user can be tricked into revealing or approvals a user can be tricked into granting.

A YubiKey can serve as one factor in a 2FA setup, or it can replace the password entirely in a passwordless FIDO2 flow. What sets it apart isn't that it's a second factor; plenty of phishable methods are also "second factors." What sets it apart is that the factor itself resists being phished in the first place, because there's no code or secret for a fake login page to capture.

What Are the Practical Limits of Using a YubiKey?

A YubiKey is a physical object, and physical objects get lost, forgotten at home, or left in a laptop bag on a business trip. Most organizations that require hardware keys issue a backup key registered to the same accounts, kept somewhere secure, precisely so one lost device doesn't lock a person out of systems they need for their job. The federal playbook referenced above makes the same point for government agencies: no single authenticator type is a complete solution, and organizations typically need a backup method alongside the primary one for account recovery and edge cases.

Cost is a real factor too, though a modest one: Yubico's own store lists its entry-level Security Key line at around $29 and its YubiKey 5 series between roughly $58 and $85, depending on form factor (Yubico Store (opens in a new tab)). Either way, it's a small line item compared to the cost of a single successful phishing breach involving AI-connected business data.

A YubiKey earns its place in an AI access strategy because it removes the single most exploited weakness in modern account takeover: the ability to trick a person into handing over a working credential. It won't fix an overly broad permissions model or a careless approval process on its own. But it closes the front door in a way passwords and app-based codes can't, and for a system built to read across everything a business knows, that's the door worth locking first.

ForgeAI / NEXT STEP

Keep the boundary where you can see it.

Book a ForgeAI demo

Sources and further reading