AI model security risks have moved from a theoretical concern to an official classification. OpenAI has publicly rated its newest frontier models as carrying “High” cybersecurity and biological risk, a disclosure that reshapes how organizations must think about deploying advanced artificial intelligence. Here is what happened and what it means.
The Official Disclosure
OpenAI made its risk assessments public this week. The language is unusually direct. According to OpenAI’s release notes, GPT-5.6 Sol, Terra, and Luna have been rated High capability in both Cybersecurity and Biological and Chemical risk, with tailored safeguards implemented to minimize associated risks.
The company coordinated with the US government before the launch. At the government’s request, access began with a limited preview for a small group of trusted partners, before wider release. None of the three models reach the threshold for High in AI Self-Improvement, which the company described as a meaningful distinction.
Why “High” Risk Classifications Matter
An official High rating is not just a label. It has real consequences for organizations that deploy these models. Specifically, it signals that these systems can meaningfully assist in activities that were previously beyond the reach of most bad actors.
Earlier in 2026, Anthropic disclosed that its most advanced models could discover vulnerabilities, build exploit chains, and assist attacks on weak enterprise networks. OpenAI’s disclosure moves in the same direction. The frontier AI models of 2026 represent a genuine uplift to anyone attempting offensive cyber or biological activities.
The Implications for Defenders
For security teams, the disclosures change the threat model. Attackers no longer need deep expertise to leverage AI for offensive purposes. This raises the floor of what is possible for a motivated but less-skilled adversary.
The practical consequence is that organizations cannot assume attackers will be limited by their own knowledge gaps. AI can fill those gaps. Therefore, defenders need to assume more capable, faster-moving adversaries and design security postures accordingly.
The Role of Confidential Computing
One response to AI model security risks is to protect the data that AI systems access rather than simply trusting the models themselves. Confidential computing keeps data encrypted even while it is being processed, which limits what any AI system, whether deployed legitimately or as part of an attack, can expose.
This data-first approach provides a layer of resilience that does not depend on predicting which model will be used or how. If the data itself is protected, the risk from a compromised or misused AI system is substantially contained.
What Organizations Should Do
The disclosures call for concrete action on several fronts. First, review how advanced AI models are used within your organization and what data they can access. Second, apply least-privilege principles so AI systems can only reach information they genuinely need.
Third, encrypt sensitive data so that even a model with broad capabilities cannot expose what it cannot read. Fourth, follow CISA’s guidance on AI security, which is increasingly specific about configuring AI systems to minimize misuse potential. The explicit rating of frontier AI as High risk is an honest acknowledgment of where the technology has arrived. For organizations using these models, matching that honesty with equally serious defensive postures is no longer optional.
This article covers ongoing security topics. Organizations should consult official vendor advisories and their security teams.
You may be interested in this article: 2026 Cyberattacks: The worst Breaches and what they have in common.
