Google Favors General-Purpose Gemini Over Specialized Cybersecurity Models

Google Cloud’s chief operating officer has confirmed that the company will not develop a dedicated cybersecurity-focused frontier model, diverging from competitors like Anthropic and OpenAI. The strategy instead hinges on general-purpose AI models, with Gemini 3.1 Pro positioned as sufficiently capable across domains, including security.

Speaking at Google Cloud Next 26, Francis DeSouza, COO of Google Cloud, explained that early industry assumptions about the need for many specialized frontier models have proven inaccurate. “What we found over time was that the core model was doing really well and that it started to get good across all domains,” he said. He pointed to coding as an example. “Coding is now done incredibly well by Gemini and you don’t need a coding specific Gemini model. We are finding that inside our security too, that models themselves are getting better and better. I believe that Gemini is a terrific model for our security. You shouldn’t expect to see a cyber version that’s different.”

DeSouza argued that the practical path forward involves combining a high-quality generalist frontier model with the right tooling and governance, rather than fragmenting effort into niche models. Google intends to pair the latest Gemini iterations with agent and platform capabilities to address cyber defense needs. Enterprises, he said, should focus on integrating strong general models into security workflows, training them with organizational context, wrapping them with access controls, and embedding them in automated detection, triage, and response pipelines.

Yinon Costica, co-founder and VP of product at Wiz, now part of Google Cloud, reinforced this view. He noted that cyber defenders possess richer, more organization-specific context than attackers, and feeding that context into a strong general model produces better defensive outcomes.

This approach contrasts with moves from competitors. Anthropic recently unveiled Project Glasswing, a cybersecurity initiative built around its Claude Mythos frontier AI model, which is fine-tuned for vulnerability detection, incident response, and adversarial reasoning. Anthropic emphasizes that domain-specific optimizations are necessary for cybersecurity’s unique challenges, such as real-time attack pattern recognition and compliance nuance. The model has been released to a select group of technology firms, including Google. A preview of Claude Mythos is also available to a select group of Google Cloud customers on Vertex AI as part of Project Glasswing.

OpenAI has taken a similar specialized route with GPT-5.4-Cyber, a variant of its flagship model tailored for defensive use cases. This launch accompanies an expanded Trusted Access Cyber (TAC) program, which provides enterprises with curated datasets, red-teaming tools, and governance frameworks to integrate the model into security operations.

Google’s decision to forgo a dedicated cybersecurity model underscores its belief that the most effective defense comes from a powerful generalist foundation, not a fragmented array of niche tools.