What We Know About GPT-5.4-Cyber So Far

Unlike Mythos Preview, GPT-5.4-Cyber launched with very little information about its capabilities. While Anthropic detailed Mythos’s ability to conduct autonomous attacks and discover thousands of vulnerabilities, OpenAI hasn’t said much about what its new model can do beyond it being more cyber-permissive than GPT-5.4.

There are significant questions as to whether this model offers new defensive capabilities for organizations or if it’s a PR pitch in response to Mythos.

“I think the takeaway might be that cybersecurity might become a new differentiator between these frontier models," Randall said before the release of the second blog post, “But at the moment, there’s more evidence that Anthropic’s is a bit more grounded.”

Randall also notes that OpenAI’s decision to open up its TAC program to a wider range of individuals and organizations than Project Glasswing also comes across as an indicator of this being a PR-driven model release.

Why GPT-5.4-Cyber Matters

For now, neither GPT-5.4-Cyber nor Claude Mythos are available for general release, but it is likely that similar models will emerge in the near future. After all, not only are Anthropic and OpenAI facing tremendous pressure to deliver new releases to justify their valuations, but there is also the prospect of other vendors developing equivalent capabilities.

Both of these limited releases indicate frontier AI models will play a bigger role in offensive and defensive operations going forward. For instance, whereas Anthropic stressed the potential offensive capabilities of Mythos in discovering vulnerabilities, OpenAI has opted to present GPT-5.4-Cyber as a defensive tool for organizations.

“GPT-5.4-Cyber reflects how quickly AI-driven vulnerability discovery is maturing. What’s notable is not just that these models can find issues, but they’re doing it more consistently and with less human effort,” Melissa Bischoping, senior director of security and product design research at security automation provider Tanium, told me via email.

“The biggest implication is speed and scale. The cat-and-mouse game we’ve played in security for years is just operating on an amplified scale now,” Bischoping said. “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 said.

The Road To Secure By Design

Beyond introducing new threats, these models have the potential to shift defensive operations as they continue to improve. "This model is just a lot better at security-related tasks. And I say a lot — it’s not really a lot. It looks like, you know, 10 to 15% better. At least, that’s what I’ve seen for Mythos,” Jeff Williams, founder and CTO of Contrast Security, told me in a video interview about GPT-5.4-Cyber.

While Williams didn’t think there would be a substantive difference between the two models, he noted that in the short term, he expects to see a glut of vulnerabilities. That being said, he’s optimistic about the ability of these tools to redefine cybersecurity in the long term.

“These models should be the key to doing security of tomorrow. But that’s not how they’re pitching them right now. And I think that’s kind of the danger, is that people are looking at these as like, ‘Ooh, we’re going to use them to find vulnerabilities and exploit them,’ and just, that’s security of yesterday," Williams said. "I’m optimistic that we can use these models to do something to actually change the industry to make things more secure.”

For Williams, the security of tomorrow is secure by design. He suggests defenders will progress from vulnerability scanning to using AI to remediate vulnerabilities and then, finally, implementing the security architecture and design to prevent these vulnerabilities from entering the code in the first place.

In any case, the fact that both Anthropic and OpenAI are enlisting support from external researchers is a recognition that it’s going to take an industry-wide effort to leverage this technology to its full potential. It’s not just a matter of developing the tools, but also helping defenders to learn how to use them effectively.