OpenAI has cut off multiple cybersecurity researchers from a program that gave them looser guardrails to test its models.
The researchers say they lost access to OpenAI's Trusted Access for Cyber, or TAC, program without warning. TAC is built for vetted security professionals and offers models with fewer of the safety restrictions applied to the general public. The idea is that legitimate researchers need to probe a model's weaknesses the way an attacker would, which is hard to do if the model keeps refusing to cooperate. It is not yet clear why access was revoked or how many researchers were affected.
This is the tension every AI lab runs into once it builds a fenced-off version of its model for security work: the same reduced guardrails that make a model useful for research also make it a bigger liability if the wrong person gets in. Losing access without explanation suggests OpenAI is erring toward caution, but it also undercuts the pitch that vetted researchers can rely on the program to do their jobs.
A cyber-testing program that researchers cannot count on having access to is worth less than one that does not exist, because it invites people to build workflows around a tool that might vanish overnight.