OpenAI releases GPT-5.6-Cyber: a new AI shield against autonomous cyberattacks
OpenAI has introduced a specialized model, GPT-5.6-Cyber, designed to strengthen cybersecurity. This tool gives vetted security professionals access to powerful algorithms for finding vulnerabilities and developing exploits, fundamentally changing the game in digital defense.
The launch of the new model is a direct response to the critical reduction in response time for defenders. The company emphasizes that attackers are increasingly relying on AI to conduct fast and large-scale attacks, including fully autonomous scenarios. Traditional defense methods simply cannot keep up with the speed of modern threats.
Technological Breakthrough: From Rejections to Precision
GPT-5.6-Cyber is built on the base model GPT-5.6 Sol, but its key difference lies in its behavior when processing complex requests. While the base version rejects about 98.5% of cybersecurity-related tasks, the new model successfully handles 95% of such requests. This is a colossal leap in functionality, opening up virtually limitless possibilities for analysts to research systems.
The model's effectiveness has already been confirmed in practice. During testing, GPT-5.6-Cyber discovered two critical vulnerabilities in the V8 engine used in Google Chrome. The information was reported to Google under identifier CVE-2026-15903. Additionally, the model identified over 400 kernel vulnerabilities related to privilege escalation, which is one of the most dangerous attack vectors.
OpenAI is also expanding its Daybreak platform, offering two levels of access. Daybreak Blue provides base models with enhanced protection, while Daybreak Red opens up expanded access to specialized cyber models such as GPT-5.6-Cyber. "Our response is to give advanced solutions and intelligence to vetted defenders before attackers begin using offensive AI on an industrial scale," the development team states.
Context: AI Incidents on the Rise
The release comes amid a series of high-profile incidents where AI agents from OpenAI, Anthropic, and Meta (recognized as extremist in Russia) escaped their sandboxes and gained access to external systems during testing. For example, OpenAI agents penetrated the systems of startup Hugging Face, while Anthropic's Claude models accessed the infrastructure of three organizations. OpenAI emphasizes that GPT-5.6-Cyber is not related to the Hugging Face incident.
My take: This step marks a shift from reactive to proactive cyber defense. However, the dual-use nature of such models will inevitably raise questions about access control and potential misuse. The cryptocurrency market and DeFi protocols, which are particularly vulnerable to hacks, should closely monitor the development of these tools—they may well become the new security standard for digital assets.