On April 30, 2026, OpenAI introduced Advanced Account Security (AAS), an opt-in setting for ChatGPT and Codex accounts that disables password-based login entirely. If you enable it, you authenticate exclusively with passkeys or FIDO-compatible hardware security keys. Email and SMS recovery are turned off. OpenAI Support cannot help you recover your account. If you lose your keys, you lose your access.
This is not a incremental security update. It is a structural replacement of the authentication layer, and understanding why OpenAI made this choice, what it means technically, and what it costs, reveals something important about where consumer-grade security is heading.
Why Passwords Fail, Even With MFA
The standard account security model looks like this: a password plus a second factor, usually a code sent via SMS or generated by an authenticator app. This model has been the baseline for over a decade. It is also the model that most account takeovers exploit.
The problem is not the password itself. The problem is that passwords are shared secrets. They can be phished, keylogged, leaked in breaches, or simply guessed. Multi-factor authentication was supposed to mitigate this by requiring something you have (your phone) in addition to something you know (your password). But the second factor is also a shared secret in many implementations: SMS codes can be intercepted through SIM swapping, and authenticator app codes can be phished through real-time proxy attacks where the attacker relays the code to the legitimate service within the time window.
Phishing-resistant authentication solves this by eliminating shared secrets entirely. Instead of proving you know a password, you prove possession of a cryptographic key. The proof is generated through a challenge-response protocol that is bound to the specific website or service. Even if an attacker tricks you into visiting a fake login page, the cryptographic proof generated by your authenticator will be bound to the wrong domain, and the legitimate service will reject it.
This is the foundation of the FIDO2 and WebAuthn specifications, which underpin both passkeys and hardware security keys. Google, Microsoft, GitHub, and others have adopted these standards. OpenAI is now aligning with this pattern.
How Advanced Account Security Works
AAS makes four structural changes to how a ChatGPT account is secured. Each change addresses a specific attack vector.
Password-based login is disabled. Once enrolled, you cannot sign in with a password. Authentication requires either a passkey (a cryptographic credential stored on your device, synced through your cloud account) or a FIDO-compatible hardware security key (a physical device like a YubiKey that generates cryptographic proofs). Both methods are phishing-resistant by design: the authentication ceremony is bound to the specific domain and cannot be replayed on a different site.
Email and SMS recovery are disabled. This closes one of the most commonly exploited attack paths. If an attacker compromises your email account or intercepts your SMS, they can use account recovery to gain access even without your password. By removing these recovery channels, AAS eliminates this vector. Recovery is limited to backup passkeys, backup security keys, and recovery keys provided during enrollment.
Session windows are shortened. Active sessions expire faster, reducing the window of exposure if a device is compromised or a session token is stolen. This is a defense-in-depth measure: even if an attacker obtains a valid session, it has a shorter useful lifetime.
Conversations from AAS-enabled accounts are excluded from model training. This is a privacy enhancement that aligns with the security posture. Users who opt into the highest security tier are likely handling sensitive personal or professional material. Excluding their conversations from training data provides an additional layer of assurance, without requiring the user to manually manage the opt-out setting.
The setting applies to both ChatGPT and Codex accounts accessed through the same login. A single enrollment covers both products.
The Enrollment Process
Enrolling in AAS requires at least two secure sign-in methods, including at least one that works across devices. Valid configurations include: a passkey and a hardware security key, two passkeys (if at least one works across devices), or two hardware security keys.
During enrollment, the system generates recovery keys. These are one-time-use codes that can restore access if you lose all other sign-in methods. OpenAI explicitly warns users to store these keys securely before completing enrollment, because once AAS is active, there is no fallback. The enrollment flow signs you out of all active sessions, ensuring that any previously compromised sessions are terminated.
The cross-device requirement is important. A passkey saved only on a single phone would lock you out if that phone is lost. Requiring at least one method that works across devices (like a hardware key or a cloud-synced passkey) ensures that losing a single device does not result in permanent account loss.
The Yubico Partnership
OpenAI has partnered with Yubico to offer preferred pricing on a bundle of two YubiKeys: the YubiKey C Nano, designed to remain plugged into a laptop USB port for daily low-friction authentication, and the YubiKey C NFC for backup and cross-device use on both laptops and mobile devices.
The partnership is a practical acknowledgment of a real barrier to FIDO adoption. Most consumers have never purchased a hardware security key. By bundling two keys at a discount and positioning one for daily use and one for backup, OpenAI and Yubico are addressing the most common failure mode of hardware-based authentication: losing the single key and being permanently locked out.
The hardware key route is optional. You can use any FIDO-compatible security key, or rely entirely on passkeys. The bundle is available to all users, not just those who enroll in AAS.
Trusted Access for Cyber: When Security Becomes Mandatory
AAS is currently opt-in for consumer accounts. But for one category of users, it will become mandatory: participants in OpenAI's Trusted Access for Cyber program, who access the company's most capable and permissive cyber models, will be required to enable AAS beginning June 1, 2026.
Organizations with Trusted Access can alternatively attest that they have phishing-resistant authentication as part of their single sign-on workflow. This is a tiered approach: individual users must prove their identity through AAS, while enterprise users can rely on organizational identity infrastructure that provides equivalent guarantees.
This mandatory requirement signals something important about OpenAI's security architecture. The most capable AI models, particularly those designed for cybersecurity work, are also the most dangerous if accessed by unauthorized parties. Requiring phishing-resistant authentication for these models is not just a user protection measure. It is an access control for frontier capabilities.
The connection to OpenAI's broader cybersecurity action plan is explicit. AAS is positioned as part of the company's effort to "broaden access to the technologies that can help protect communities, critical systems, and our national security." The framing aligns identity assurance with capability access: the stronger your authentication, the more powerful the tools you can access.
The Tradeoff: Key Management as a First-Class Responsibility
AAS is not a free upgrade. It transfers the burden of account recovery from OpenAI to the user, and this transfer has consequences.
Under standard authentication, losing your password is recoverable: you can reset it via email, verify your identity with OpenAI Support, or use a backup code. Under AAS, the recovery surface shrinks to backup keys and recovery codes. If you lose both your primary and backup authentication methods, and you have not stored your recovery keys, your account is gone. OpenAI Support will not be able to help.
This tradeoff is intentional. It reflects a security engineering principle: the most reliable way to prevent unauthorized recovery is to make recovery itself impossible without cryptographic proof of identity. The cost is that legitimate recovery also becomes harder. This is the same tradeoff that hardware wallet users in cryptocurrency accept, and that enterprise security teams enforce with privileged access management systems.
For most users, the risk calculus is straightforward. If your ChatGPT account contains sensitive conversations, proprietary code, or professional data, the protection against phishing-based account takeover likely outweighs the risk of losing access due to key mismanagement. If your account contains nothing that would be catastrophic to lose, the standard security model is sufficient.
The practical recommendation is simple: if you enroll, buy two hardware keys. Keep one plugged into your primary device. Store the other in a safe place. Save your recovery keys in a password manager or a physical safe. This three-layer backup strategy eliminates virtually all risk of permanent lockout while maintaining the full phishing resistance that AAS provides.
Technical Context: Why This Matters Beyond OpenAI
OpenAI's AAS implementation is not technically novel. Passkeys and FIDO2 hardware keys are industry standards. Google, Microsoft, Apple, and GitHub have all implemented similar systems. What makes AAS noteworthy is the context in which it is being deployed.
ChatGPT is used by over 400 million people. Many of these users have never encountered phishing-resistant authentication. By making it available (and eventually mandatory for high-privilege access), OpenAI is bringing a security technology that has been confined to enterprise environments to consumer scale.
The conversation exclusion from model training is also significant. It creates a security-privacy alignment: users who take the strongest security measures receive the strongest privacy protections. This is a design pattern that other services may adopt, as it incentivizes security adoption through privacy benefits rather than through fear of compromise.
For security practitioners, AAS is also a case study in how to design an opt-in security system that is strict enough to provide real protection while remaining usable. The requirement for two cross-device methods, the YubiKey bundle, and the explicit warnings about recovery are all design choices that balance security and usability at the boundary where most systems fail.
FAQ
What is OpenAI Advanced Account Security?
Advanced Account Security (AAS) is an opt-in security setting for ChatGPT and Codex accounts that replaces password-based login with phishing-resistant authentication using passkeys or FIDO-compatible hardware security keys. It also disables email and SMS account recovery, shortens session windows, and excludes conversations from model training.
How do passkeys differ from passwords?
Passkeys are cryptographic credentials stored on your device and synced through your cloud account. Unlike passwords, they are not shared secrets. Authentication uses a challenge-response protocol bound to the specific website, making it resistant to phishing, keylogging, and credential replay attacks.
What happens if I lose my security keys with AAS enabled?
If you lose all sign-in methods, you can use the recovery keys provided during enrollment. If you also lose your recovery keys, your account cannot be recovered. OpenAI Support cannot assist with recovery for AAS-enrolled accounts. This is why AAS requires at least two sign-in methods, including one that works across devices.
Who needs to use Advanced Account Security?
AAS is currently optional for all eligible personal ChatGPT accounts. Starting June 1, 2026, it becomes mandatory for individual participants in OpenAI's Trusted Access for Cyber program, who access the company's most capable cybersecurity models. Organizations can alternatively attest to having phishing-resistant authentication in their SSO workflow.
Does Advanced Account Security work with Codex?
Yes. AAS applies to both ChatGPT and Codex accounts accessed through the same login. A single enrollment covers both products.
References
- OpenAI, "Introducing Advanced Account Security," April 30, 2026. https://openai.com/index/advanced-account-security/
- OpenAI Help Center, "Advanced Account Security." https://help.openai.com/en/articles/20001221-advanced-account-security
- Help Net Security, "ChatGPT Advanced Account Security Adds Passkeys and Hardware Keys," May 4, 2026. https://www.helpnetsecurity.com/2026/05/04/openai-chatgpt-advanced-account-security/
- OpenAI, "Cybersecurity in the Intelligence Age" (Cybersecurity Action Plan). https://cdn.openai.com/pdf/7ca95dce-4424-4b62-9eab-89233bb38f82/oai-cybersecurity-action-plan.pdf