OpenAI Warns of Escalating AI Cybersecurity Threats as Models Advance

OpenAI has issued a stark warning about the cybersecurity risks posed by its next-generation AI models, citing their potential to develop zero-day exploits and facilitate complex intrusion operations. The company revealed that its GPT-5.1-Codex-Max model achieved a 76% success rate in capture-the-flag challenges, a significant leap from GPT-5's 27% performance just months prior.

In response to these risks, OpenAI is rolling out a multi-layered defense strategy. The measures include stringent access controls, hardened infrastructure, and advanced monitoring systems. Aardvark, an AI security agent currently in private beta, will scan code for vulnerabilities and suggest patches proactively.

The company is establishing a Frontier Risk Council comprising external cybersecurity experts to address emerging threats. Tiered access programs will be implemented for cyberdefense users, ensuring appropriate safeguards accompany advanced capabilities.