OpenAI is scaling its Trusted Access for Cyber (TAC) cybersecurity program to thousands of individual defenders and teams responsible for protecting critical software.

The specialized model is designed to give cybersecurity teams a technical edge over AI-augmented attackers and marks a strategic shift toward "cyber-permissive" AI. This means that safety guardrails are calibrated to allow security professionals to perform sensitive tasks, such as reverse engineering malware, that standard AI models might typically refuse.

“For years, we’ve been building a cyber defense program on the principles of democratized access, iterative deployment, and ecosystem resilience,” the company said. “In preparation for increasingly more capable models from OpenAI over the next few months, we are fine-tuning our models specifically to enable defensive cybersecurity use cases, starting today with a variant of GPT‑5.4 trained to be cyber-permissive: GPT‑5.4‑Cyber.”

New tier of defensive AI

That new variant is touted as a fine-tuned version of the company's latest architecture, specifically trained to assist in advanced defensive workflows.

Some of the capabilities of the new model include binary reverse engineering to analyze compiled software for vulnerabilities and malware without requiring access to the original source code and reduced refusal boundaries. This would allow verified researchers to query the model on dual-use topics that would trigger safety blocks in consumer versions.

The model has also incorporated agentic patching, an integration with Codex Security to automatically monitor codebases and propose fixes in real-time.

OpenAI is also introducing a multitiered verification system that uses automated identity checks for individuals and deeper vetting for enterprise teams. This framework allows legitimate defenders to access more permissive capabilities while maintaining accountability through tiered restrictions and visibility measures for high-risk use cases.

3,000 critical and high-severity vulnerability fixes

The release comes as OpenAI acknowledges that cybersecurity risks are "already here and accelerating." The company reported that its Codex Security system has contributed to more than 3,000 critical and high-severity vulnerability fixes since its private beta launch.

"Our goal is to make these tools as widely available as possible while preventing misuse," the company stated, emphasizing a shift toward democratized access for legitimate actors.

By moving security from episodic audits to ongoing risk reduction directly in the developer workflow, OpenAI aims to ensure that as AI models become more capable, the systems protecting the world’s digital infrastructure scale at the same velocity.

The company said that in the future, more powerful models will require more expansive and agentic defenses to ensure that AI-driven defense consistently outpaces AI-driven threats.

The release of OpenAI’s GPT-5.4-Cyber comes a week after Anthropic’s Claude Mythos announcement, a new general-purpose language model for security tasks, demonstrating a further divide in how the rivals handle frontier cybersecurity capabilities.

Melissa Bischoping, senior director of security and product design research at Tanium, noted models such as Mythos and GPT-5.4-Cyber reflect a quick evolution in AI-driven vulnerability discovery.

"What’s notable is not just that these models can find issues, but that they’re doing it more consistently and with less human effort," Bischoping explained. "There were also a very small number of individuals globally who are deep experts in this domain. These advances start to remove the previous bottleneck – human time and expertise – in exploit development and allow the experts to focus on the highest value bugs more efficiently."

For Bischoping, programs like OpenAI's TAC program are a step in the right direction, even if they aren't ideal remedies. 

"We’re seeing multiple AI companies invite industry in to test or review their emerging products, including Anthropic with its limited release of Mythos," Bischoping noted. "The mechanics between the programs are different, and I don’t think either are a perfect solution, but the spirit and intent appears to be in the right place for both. Frontier labs are investing in collaborative defense. I hope these programs reinforce that this is becoming the standard for how these models are introduced responsibly."

Bischoping added that competitive dynamics between AI vendors and the advancements in AI in general are accelerating timelines. According to the Tanium exec, models like Mythos compete at ranges and evaluations, such as CyberGym, that they hadn't been able to perform before. 

"While GPT-5.4 hasn’t posted any such results for independent eval, it’s OpenAI’s answer to that level of capability," Bischoping said. "What we’re seeing with OpenAI, alongside moves from others in the space, is that these models are improving quickly and that level of capability won’t stay contained to a small group for long. Anthropic is estimating open-weight models at parity capability within six to 18 months, and I think it will be closer to six than 18.

"The biggest implication is speed and scale," Bischoping continued. "The cat-and-mouse game we've played in security for years is just operating on an amplified scale now. Historically, the disclosure-to-exploitation window was around 63 days. That shrank to as few as five days in 2024. This new class of model has compressed that to potential hours. As models get better at identifying vulnerabilities, the window between discovery and exploitation continues to shrink. That puts pressure on organizations to move toward continuous validation and real-time response. Open-weight models hitting the ecosystem will increase that pressure."

Bischoping doubts organizational change control can move fast enough to act on what these models discovers before they’re out in the wild, as "few can prove they’ve fixed them across every endpoint."

"This is amplified when you consider that the National Vulnerability Database (NVD) is already struggling to keep up, with nearly half of all vulnerabilities added in the past year perpetually stuck ‘awaiting analysis.’ That gap between discovery, analysis, and verified remediation is where exposure lives in this new dynamic," Bischoping noted.