在 2026 年 4 月 30 日 13:30（原文时间为 1:30 PM），OpenAI 公开表示，为了降低 ChatGPT 与 Codex 帐户成为攻击目标的风险，将推出可选的 Advanced Account Security，作为额外的一层帐户防护。这项机制的设计目标是让帐户被接管（Account takeover）更难发生。对应的概念并不新颖：Google 的 Advanced Protection 已运行近十年。然而 OpenAI 指出，随著人工智慧服务快速全球化，基础安全架构有更迫切的整备需求，而此举是其本月早前提出之更广泛资安策略的延伸。
OpenAI 将有高风险需求的使用者，像是记者、民选官员、政治异议者、研究人员与高度安全意识者，视为主要受保护对象之一。启用后，使用者不能再使用一般密码，必须改以至少两把实体安全金钥（physical security keys）或 passkeys 登入，才能显著降低钓鱼攻击成功率。此模式也移除了以电子邮件与简讯为主的帐户复原路径，改由复原金钥、备援 passkeys 或实体金钥处理。OpenAI 已与 Yubico 合作，对 Advanced Account Security 用户提供较低价的 YubiKey 套组。
此外，启用者一旦遇到帐户问题，OpenAI 支援团队将无法协助复原，因为复原选项不再由支援系统掌控，以阻断攻击者针对客服入口进行社交工程。系统还会缩短签入视窗与工作阶段持续时间，并在有登入活动时产生警示，导向仪表板检视活跃的 ChatGPT 与 Codex 会话。另一方面，虽然所有使用者都可选择将对话排除于模型训练外，Advanced Account Security 用户在此设定上为预设启用。OpenAI 的 Trusted Access for Cyber 成员需在 6 月 1 日起改用此功能，或改以企业式单一登入机制提供防钓鱼验证宣誓（attestation）。
On April 30, 2026 at 1:30 PM, OpenAI announced an optional Advanced Account Security mode for people who may be targeted by attackers on their ChatGPT or Codex accounts. The new layer is meant to make account takeover attacks much harder. This model is not a new security idea—Google’s Advanced Protection has been available for nearly a decade—but OpenAI said that as AI services scale rapidly worldwide, stronger baseline protections are needed. The company positioned this launch as part of its broader cybersecurity strategy announced earlier that month.
OpenAI said users in higher-risk roles—such as journalists, elected officials, political dissidents, researchers, and security-conscious users—can opt in. After enabling the feature, regular passwords are removed; users must use at least two physical security keys or passkeys, which is intended to greatly reduce successful phishing. Email and SMS recovery routes are also removed, replaced by recovery keys, backup passkeys, or physical keys. OpenAI also said it partnered with Yubico to offer lower-cost YubiKey bundles for Advanced Account Security users.
With this mode on, users can no longer ask OpenAI support to recover accounts, because support no longer controls recovery options; this is designed to block attackers from exploiting support channels through social engineering. Advanced Account Security also shortens sign-in windows and session lifetimes, and it alerts users when logins occur so they can review active ChatGPT and Codex sessions on the dashboard. It also sets conversation-training opt-out to default on for these users, and members of OpenAI’s Trusted Access for Cyber program are required to enable this mode from June 1 unless they submit an alternative attestation showing phishing-resistant authentication through enterprise SSO.