Access Revocations Hit a Vetted Cyber Program
OpenAI said a technical issue caused some vetted security researchers to lose access to a limited cyber program designed to relax certain safeguards for authorized defensive work.
Several researchers reported on OpenAI’s official support forums and on X that their access to the Trusted Access for Cyber program, or TAC, had suddenly stopped working. When they opened ChatGPT’s Cyber page, they saw messages saying their identity could not be verified or that their account was “ineligible at this time.”
TAC is OpenAI’s vetted access program for cybersecurity researchers. It gives approved users access to more advanced models with fewer cybersecurity restrictions than those available to regular users. The purpose is not to create a general hacking mode, but to support trusted defenders who are trying to find and report vulnerabilities so companies can patch them faster.
Why These Programs Exist
Cybersecurity research often requires asking AI systems about code flaws, malware behavior, exploit validation, or incident response. Those same capabilities can also be useful to criminals. That is why OpenAI requires TAC applicants to submit identification and go through vetting before they are approved.
Anthropic operates a comparable system called the Cyber Verification Program, or CVP. Both programs reflect the same balancing act: give legitimate researchers stronger tools while keeping cybercriminals and malicious hackers away from models that could help them find bugs or build exploits.
The affected access appears to involve Daybreak Blue, OpenAI’s latest vetted tier for individual researchers. OpenAI launched Daybreak Blue on August 10 and describes it as providing access to “frontier general-purpose models,” including GPT‑5.6 Sol, with safeguards tailored for authorized defensive security work. The company says it is the recommended starting point for most defenders and supports vulnerability discovery, secure code review, malware analysis, incident response, and patch validation.
What OpenAI and Researchers Said
It is still unclear how many people were affected. TechCrunch spoke with five researchers who said they had experienced the problem. All five said they lived outside the United States and Europe, suggesting the issue may have been limited to certain regions, though OpenAI has not confirmed that as the cause.
One researcher shared an email from OpenAI saying their Daybreak Blue access had been revoked “due to a technical issue affecting a limited number of users.” The message added that the issue was on OpenAI’s end and was “not the user experience” the company wanted to deliver.
A researcher in an OpenAI forum thread said the company’s support team also referred to a “recent technical issue” that caused some users to lose Daybreak Blue access. In both cases, OpenAI told affected researchers to reapply and complete the verification process again.
OpenAI also pointed TechCrunch to a post saying a limited set of users’ Daybreak Blue access was no longer active and that they would need to re-verify to maintain access.
A Wider Debate Over AI Guardrails
OpenAI introduced Daybreak Blue alongside a higher tier called Daybreak Red. According to the company, Daybreak Red gives vetted users access to models built specifically for cybersecurity research and supports authorized vulnerability research, exploit validation, and security testing.
The incident comes amid broader complaints from both defensive and offensive security researchers about AI guardrails. In this context, guardrails are restrictions or refusal mechanisms that stop models from assisting with risky requests. Researchers argue that overly broad restrictions can interfere with legitimate work, while AI companies must prevent the same systems from being used for abuse.
The episode shows how difficult that middle ground is becoming. Vetted access programs may be one of the more practical ways to support legitimate security research, but they depend on reliable identity checks, stable permissions, and clear recovery processes when something goes wrong. As advanced models become more relevant to vulnerability discovery and incident response, the operational quality of these access systems will matter almost as much as the models themselves.