研究人员称 OpenAI 撤销其有限网络安全项目访问权限

多名安全研究人员表示，OpenAI 突然撤销了他们对一项有限访问计划的权限。该计划移除了其 AI 工具在网络安全研究用途上的部分限制。OpenAI 确认，该问题由一项错误导致。

周三，多名研究人员在 OpenAI 官方支持论坛和 X 平台上表示，他们对 Trusted Access for Cyber（TAC）项目的访问权限已被撤销。这些人士称，当他们打开 ChatGPT 的 Cyber 页面时，页面显示其身份无法得到验证，或其账户“目前不符合资格”。

TAC 是一项特殊计划，OpenAI 通过该计划向经过审核的研究人员提供公司最先进 AI 模型的访问权限，其网络安全护栏少于普通用户可访问模型的限制。Anthropic 也提供类似项目，即 Cyber Verification Program（CVP，网络安全验证计划）。

TAC 和 CVP 的理念是向可信赖的防御方提供更强大的模型，使他们能够向企业报告漏洞与安全弱点，从而推动缺陷更快得到修复。其目标也是防止网络犯罪分子和恶意黑客获得这些模型，并利用它们寻找漏洞、开发攻击利用程序以入侵企业。

要获得 TAC 的访问权限，网络安全研究人员必须提交身份证明并通过 OpenAI 审核。

目前尚不完全清楚这些研究人员为何被撤销 TAC 访问权限，也不清楚有多少人受到了这一问题的影响。

TechCrunch 与五名表示遇到该问题的研究人员进行了交流。其中一名研究人员称，OpenAI 发送了一封电子邮件，表示其对 Daybreak Blue——TAC 最新的审核访问层级——的权限已被撤销，原因是“影响少量用户的技术问题”。

“这是我们这边的问题，并非我们希望向用户提供的体验，”该消息写道。这名研究人员将邮件分享给了 TechCrunch。

在 OpenAI 论坛的一个讨论帖中，一名研究人员写道，在其联系官方支持团队后，公司也提及一项“近期技术问题”，导致部分用户失去 Daybreak Blue 的访问权限。

在这两则消息中，OpenAI 都要求研究人员重新申请并完成验证流程。

所有与 TechCrunch 交流的研究人员都表示，他们居住在美国和欧洲以外，这意味着此次权限撤销可能仅限于某些地区。

OpenAI 向 TechCrunch 提供了一条推文，其中该公司表示，“一小部分用户对 Daybreak Blue 的访问权限已不再有效，他们需要重新验证以维持访问权限。”

Daybreak Blue 是面向个人研究人员的最新层级。据 OpenAI 介绍，该层级于 8 月 10 日推出，提供对“前沿通用模型，包括 GPT‑5.6 Sol”的访问，并配备“针对获授权防御性安全工作而定制的保障措施”。“它是大多数防御人员推荐的起点，支持漏洞发现、安全代码审查、恶意软件分析、事件响应和补丁验证。”

与此同时，该公司还推出了名为 Daybreak Red 的更高级别层级，提供专为网络安全研究打造的模型访问权限，允许经过审核的用户开展“获授权的漏洞研究、利用验证和安全测试”。

近几个月来，防御性和进攻性安全研究人员都曾抱怨 Anthropic 和 OpenAI 设置的护栏，认为这些护栏妨碍了他们开展合法工作。

已根据 OpenAI 的回应更新。

Several security researchers say OpenAI suddenly revoked their access to a limited-access program that removes some restrictions on using its AI tools for cybersecurity research. OpenAI confirmed that the issue was caused by an error.

On Wednesday, multiple researchers on OpenAI’s official support forums and on X reported having their access to the Trusted Access for Cyber (TAC) program revoked. The people said that when they opened ChatGPT’s Cyber page, a message appeared saying their identity could not be verified or that their account “is ineligible at this time.”

TAC is a special program through which OpenAI offers vetted researchers access to the company’s most advanced AI models with fewer cybersecurity guardrails than the models that regular users can access. Anthropic offers a similar program called the Cyber Verification Program, or CVP.

The idea behind TAC and CVP is to give trusted defenders better models so they can report bugs and vulnerabilities to companies, with the aim of getting flaws patched faster. The goal is also to prevent cybercriminals and malicious hackers from accessing those models and use them to find bugs and develop exploits to hack companies.

To get access to TAC, cybersecurity researchers have to submit an ID and be vetted by OpenAI.

At this point, it’s not entirely clear why these researchers are getting their access to TAC revoked, nor how many people have had this issue.

TechCrunch spoke to five researchers who said they have had this problem. One researcher said OpenAI sent an email saying that their access to Daybreak Blue, the latest vetted tier of TAC, was revoked “due to a technical issue affecting a limited number of users.”

“This was an issue on our end, and not the user experience we want to deliver,” read the message, which the researcher shared with TechCrunch.

In a thread on OpenAI’s forums, a researcher wrote that after they reached out to the official support, the company also cited a “recent technical issue” that caused some users to lose access to Daybreak Blue.

In both messages, OpenAI asked the researchers to reapply and complete the verification process.

All the researchers TechCrunch spoke to said they live outside of the U.S. and Europe, suggesting the revocations may be limited to certain regions.

OpenAI referred TechCrunch to a tweet in which the company said a “limited set of users’ access to Daybreak Blue is no longer active and they will need to re-verify to maintain their access.”

Daybreak Blue is the latest tier for individual researchers that grants access to “frontier general-purpose models, including GPT‑5.6 Sol, with safeguards tailored to authorized defensive security work,” according to OpenAI, which launched it on August 10. “It is the recommended starting point for most defenders, supporting vulnerability discovery, secure code review, malware analysis, incident response, and patch validation.”

At the same time, the company also introduced a higher tier called Daybreak Red that gives access to models made specifically for cybersecurity research, which allow vetted users to do “authorized vulnerability research, exploit validation, and security testing.”

In recent months, both defensive and offensive security researchers have complained about the guardrails imposed by Anthropic and OpenAI, arguing that the guardrails prevent them from doing legitimate work.

Updated with comment from OpenAI.