AI-powered cybersecurity is poised for a major surge.
On Tuesday, OpenAI announced that it was scaling up its Trusted Access for Cyber (TAC) program to thousands of “verified individual defenders” and hundreds of teams involved in protecting critical software. These users will have access to a GPT-5.4 variant called GPT-5.4-Cyber that was trained to be “cyber-permissive”.
Specifically, the model was fine-tuned to have fewer capability restrictions and includes new capabilities for these workflows, such as binary reverse engineering, according to the post. In the coming months, OpenAI says it will fine-tune its models to better cater to cybersecurity needs.
Because of the permission updates to GPT-5.4-Cyber, OpenAI is first deploying it to vetted security vendors, organizations, and researchers who are part of the TAC's new tiers of access, with people willing to work with OpenAI to authenticate themselves as cybersecurity defenders. Interested TAC members apply for the cyber defenders pilot via this Google form.
TAC users will still be able to take advantage of the benefits introduced with the program in February, including reduced friction with safeguards on models. Becoming part of the TAC is meant to be simple: individuals apply with the form, and enterprises request trusted access for their team through their OpenAI rep.
OpenAI also re-emphasized its mission and strategy for cybersecurity-related AI solutions, with key highlights including the need to expand access while preventing misuse, to deploy iteratively, and to invest in the ecosystem.
This announcement, of course, follows Anthropic’s launch of Project Glasswing, an initiative to secure critical software in the age of AI, which included the launch of Claude Mythos, a model more advanced than its current top-tier model, Opus. The biggest difference is that Anthropic’s launch was much more limited, with only launch partners and 40 additional organizations getting access.
“Anthropic has taken a more contained route, with a clear focus on how these systems behave when pushed beyond their intended use,” said Hanah-Marie Darley, chief AI officer and co-founder of Geordie AI, an AI Governance platform. “OpenAI is opening access more broadly and placing weight on the safeguards around the model. Both approaches are reasonable.”
Our Deeper View
Anthropic's Mythos announcement dropped while I was at HumanX last week, making it a hot topic in nearly every conversation I had there. A common thread was that withholding the model was a calculated move for media coverage, investor interest, and safety-first branding. I disagree. The decision still reflects a sense of responsibility, because Anthropic had to know that competitors like OpenAI had similar models and would release them more broadly, given their different risk tolerance. Choosing restraint meant ceding ground and facilitating competition. As for OpenAI’s approach, expanding access also inevitably entails greater risk.