Anthropic & OpenAI split on cyber AI release strategy

Anthropic and OpenAI have launched new artificial intelligence models that can autonomously search for cyber vulnerabilities, but they are taking sharply different approaches to releasing them.

The launches come as security specialists warn that large language models now match or exceed human experts in several key stages of offensive security work. These models can analyse codebases, identify exploitable flaws, and generate working exploits in ways earlier systems could not reliably achieve.

Anthropic is rolling out a cybersecurity-focused version of its Claude family called Claude Mythos Preview. The model operates within a tightly controlled initiative called Project Glasswing. Participants include major technology providers such as AWS, Apple, Google, Microsoft, Nvidia and Cisco, along with security firms including CrowdStrike and industry bodies such as the Linux Foundation.

Anthropic has tied this closed approach to the model's ability to autonomously discover severe vulnerabilities, signalling that unconstrained access would not currently be safe. Mythos is not available to general customers and remains in what Anthropic describes as a restricted preview, with direct oversight of partners and projects.

OpenAI has taken a broader approach with its new system, GPT-5.4-Cyber, which sits within an expanded Trusted Access for Cyber programme. Access is still gated by verification checks, but the scheme extends availability to thousands of individual security practitioners and hundreds of corporate security teams.

OpenAI describes this as a "democratised defence" strategy, arguing that defenders need wider access to advanced tools to keep pace with threat actors already experimenting with generative AI. It is betting that a large pool of vetted users will improve security outcomes faster than a smaller, closed group.

Jonny Scott, head of cyber advisory at UK-based Phoenix Software, sees the divergence as evidence of a split in how leading AI companies view cyber risk. In his view, the ability of these models to carry out independent vulnerability discovery changes the threat landscape.

He is particularly concerned about the wider reach of OpenAI's model compared with Anthropic's, noting that GPT-5.4-Cyber is likely to end up in the hands of a much broader and more varied user base than Claude Mythos Preview.

"The AI hype is real - and the change is coming, for good and bad. But it's interesting to see these two companies taking such different views on how these solutions should be brought to market," Scott said. "Both approaches aim to strengthen defence - but they clearly have very different risk tolerances."

Scott argues that the question in cybersecurity is no longer whether AI will become central to defence strategies. The real issue, he says, is whether suppliers and customers can integrate these systems quickly while avoiding new risks created by the tools themselves.

Edward Wu, founder and chief executive of US-based security start-up Dropzone AI, also sees a sharp jump in real-world performance in the latest models. He points particularly to their impact on work that once required rare expertise within offensive security teams.

"The latest batch of LLMs, such as Claude Mythos, showed a step-function increase in both the ability to discover critical vulnerabilities in open-source projects and the ability to generate working exploits, which earlier models consistently struggled with. This is concerning because cyber attackers traditionally were constrained by expertise and cost, and models like Mythos could significantly reduce the effort and resources required to breach traditional cybersecurity perimeter defenses by discovering and weaponizing more 0-day vulnerabilities.

"While model providers such as Anthropic and OpenAI are currently restricting these models to defensive uses, it's imminent that similar capabilities will become more widely accessible to actual attackers over the next 12 to 18 months as open-weight models catch up. For defenders, this means assuming much shorter patching windows, adopting an 'assume breach' mindset, and investing in automation to operate at machine speed and scale. Improving detection and response times and accelerating vulnerability remediation processes becomes critical as prevention alone won't keep pace," Wu said.

Security teams already face mounting strain from rising incident volumes, skills shortages and legacy systems. Analysts say more advanced AI assistance could lower the barrier for overworked defenders dealing with long backlogs of vulnerabilities and alerts.

At the same time, both Scott and Wu see a credible risk that these models could leak advanced tradecraft into the wider ecosystem as their methods spread through open-source communities or criminal forums. They warn that systems trained for defence could also support offensive campaigns if access controls fail or equivalent open-weight models emerge.

Project Glasswing appears designed around that concern. Anthropic has selected a small group of partners with established internal security processes and significant engineering resources, giving it closer oversight of how Claude Mythos Preview is tested and where its outputs go.

OpenAI's broader programme follows the opposite logic. It is expanding the pool of users that can query GPT-5.4-Cyber, subject to verification and usage policies, on the assumption that putting these tools in the hands of thousands of legitimate defenders will improve collective resilience, even if it increases the potential for misuse.

Industry observers see echoes of earlier debates over vulnerability disclosure and exploit development. Some researchers prefer tightly controlled coordination with a small group of trusted partners, while others favour rapid, wide dissemination of information on flaws and mitigations.

For now, both Anthropic and OpenAI say they restrict dangerous outputs and focus use on defensive scenarios such as penetration testing, red-teaming and code review. External experts warn that these restrictions will come under pressure as open-weight competitors pursue similar performance without equivalent guardrails.

The emerging split between closed and broad release models presents customers with a stark choice in how they engage with AI-driven cybersecurity. It also sharpens scrutiny of how AI providers decide who can access systems that approach, and in some niches now exceed, advanced human expertise in cyber offence.

Scott sees that tension as the defining issue for the sector in the next phase of AI deployment. In his view, the balance between control and scale will determine whether AI strengthens digital defences or amplifies the very threats it is meant to address.