Key Highlights

– OpenAI says future models may reach ‘High’ cybersecurity capability.
– Models could help develop zero-day exploits without strict safeguards.
– New defensive tools, red teaming, and trusted access programs announced.
– Aardvark security agent enters private beta to detect vulnerabilities.
– OpenAI pushes for shared industry standards through Frontier Model Forum.

OpenAI has issued a sharp warning about its next wave of AI models. The company now expects future systems to hit ‘High’ cybersecurity capability, a level where models might help craft zero-day exploits or assist complex intrusion operations. This shift pushes OpenAI to rethink how these models should operate in the real world.

AI Models Are Getting Stronger Faster

OpenAI says cyber skills inside its models have jumped rapidly. CTF challenge scores moved from 27% on GPT-5 in August to 76% on GPT-5.1-Codex-Max in November. This rise signals a new era where AI can automate tasks once reserved for elite hackers.

Therefore, OpenAI plans as if each new model could reach the highest tier of capability. This approach forces the company to design safeguards that support defenders but block malicious actors.

OpenAI Builds a Defense-First Strategy

The company says it will combine infrastructure hardening, access controls, monitoring, and detection tools. It will also train models to reject unsafe cyber requests while still supporting learning and defensive workflows.

OpenAI has expanded system-wide monitoring to catch suspicious activity early. Unsafe prompts may be blocked or routed to safer models. Human review will handle severe cases.

Moreover, OpenAI has partnered with expert red-teaming groups. These teams test the entire system like real attackers to expose weak spots before adversaries do.

Aardvark Brings Automated Cyber Defense

OpenAI also announced Aardvark, an agentic security researcher that scans large codebases and suggests patches. It has already spotted new flaws in open-source projects. The tool is now in private beta. OpenAI plans to offer free support to select non-commercial repositories.

Trusted Access Program Coming Soon

To avoid misuse, OpenAI is preparing a trusted access program for cyberdefense users. This program will offer tiered access to stronger capabilities while keeping harmful power away from bad actors. The company is still deciding which features should require more strict controls.

Industry Collaboration To Address Growing Risks

OpenAI says no single company can solve frontier cyber risk. Therefore, it works with global partners through the Frontier Model Forum to share threat models and best practices. The goal is to ensure stronger security across the ecosystem.

OpenAI adds that this work will continue as AI grows more capable and more challenging to secure.