Who Will Win Race for AI-Driven Cyber - Anthropic or OpenAI?

Two artificial intelligence leaders - Anthropic and OpenAI - are taking sharply different approaches to introducing new models that could upend processes for finding and fixing software vulnerabilities, and analysts say the winner will be the company that does better at both.

See Also: AI Security Risks Rise With Agentic Systems

Introduced just weeks apart, Anthropic's Claude Mythos Preview is good at vulnerability discovery and exploitation, while OpenAI's GPT-5.4-Cyber is placing more emphasis on integrating AI outputs into existing security operations center workflows, including feeding vulnerability findings into security tools, enabling automated patching and supporting mitigation strategies, said Frank Dickson, group vice president for IDC's Security & Trust research practice.

Purpose-built models will deliver stronger performance within defined domains, said Jeff Pollard, vice president and principal analyst at Forrester.

"Once you start building a model to do a more narrow set of use cases, that means that it's more effective at those use cases," Pollard told ISMG. "That's a positive thing. If they have different scenarios that they're best at, that's a net positive from a cybersecurity perspective."

Anthropic has leaned heavily into blog posts, model cards and high-profile partnerships, which Pollard said helps both practitioners and non-practitioners understand what's possible even if the underlying capabilities aren't fundamentally different from competitors. OpenAI, in contrast, has been more restrained, which he said may create the impression of lagging innovation even if that's not the case (see: OpenAI Courts Banks in Trusted Access for Cyber Partner Push).

"That's less about the quality of the model or what it can do and frankly more about Anthropic's ability to amplify its message and make it flash with some of the announcements in comparison to OpenAI," Pollard said.

If OpenAI and Anthropic can produce models that exceed the security capabilities of vendor-developed systems, Pollard wondered about the value in cyber vendors building proprietary AI. IDC said the value will shift toward connecting AI-driven insights to enterprise workflows, while Pollard said cyber vendors will provide differentiated solutions by how effectively they enable organizations to use AI models in practice.

"Does it make sense for them to be trying to train their own AI when the frontier labs keep coming out with AI that's just as good as what they offer?" Pollard said. "It's going to be about the other stuff that they're great at - their core capabilities - not so much the AI stuff."

As the volume of vulnerabilities grows, Dickson said there may come a point at which continuous patching is no longer practical. In that case, organizations may choose to rewrite applications from scratch rather than maintain increasingly fragile code bases. When vulnerabilities accumulate faster than they can be addressed, the economics shift, making replacement more attractive than incremental fixes.

"It turns into whack-a-mole," Dickson told ISMG. "At a certain point, it becomes easier just to rewrite the whole application than it is to try to figure out how to patch it." If a cost-benefit analysis shows it's easier to write code and the scale of vulnerabilities grows, organizations may be "driven to inherently rewrite your code from scratch more often," Dickson said.

Larger context allows models to better understand how different components of a system interact, identify subtle vulnerabilities and even chain multiple weaknesses together into viable exploits, Pollard said. The same capability also improves defensive use cases. As models become better at understanding systems holistically, they become more effective at both breaking and securing them.

"You can point it to larger chunks of code, and as a result of that, it can ingest that code, understand it and reason about it better, which is going to help it ultimately find more issues and also generate more exploits," Pollard said.

Anthropic's Project Glasswing limits Claude Mythos Preview access to a relatively small group of high-profile partners, which Dickson said allows for testing and refinement in a constrained environment but restricts broader industry participation. OpenAI's Trusted Access for Cyber program appears more scalable and more governed, Dickson said, aiming to broaden availability while maintaining oversight.

"It seems as though the OpenAI approach is much more considered when it comes to cybersecurity," Dickson said. "It seems like OpenAI has thought about what the ramifications are of dumping a very large and powerful model into a world full of 50 years of flawed code."

Critical infrastructure systems rely on outdated and unpatched software, and Dickson said introducing AI systems capable of rapidly identifying vulnerabilities into this environment could expose weaknesses faster than organizations can remediate them. With dozens of players in critical infrastructure, limiting access to a small group risks excluding organizations that play vital roles in securing niche environments.

"The OpenAI approach will be more democratizing," Dickson said. "It's more work on their part to be able to validate and vet all of the users and companies to make sure that they are indeed legitimate. But for the security ecosystem, I'm appreciative of the effort."