OpenAI Announces Security Innovations with Official Release of GPT-5.6: Blocking Prompt Injections via Automated Red Teaming Model 'GPT-Red'

On July 9, 2026, OpenAI officially launched the GPT-5.6 family of next-generation frontier AI models after delays following a national security review request from the U.S. government. This release goes beyond simple performance improvements, featuring a systematic strengthening of defenses against prompt injection attacks by introducing an automated red-teaming model called 'GPT-Red'.

GPT-5.6 consists of three tiers—Sol, Terra, and Luna—with each model optimized for security and efficiency to meet the needs of enterprises and government agencies. In particular, the top-tier model, Sol, demonstrates overwhelming performance in cybersecurity and coding tasks, setting a new standard for AI security.

First unveiled as a limited preview on June 26, 2026, GPT-5.6 was only supplied to the full market on July 9, later than originally planned. This was due to a rigorous national security verification process following U.S. government concerns that powerful AI systems could be misused. This official release suggests the emergence of a new regulatory model where frontier AI systems undergo strict government scrutiny before broad commercial deployment.

The release of GPT-5.6 is not only a technological advancement but also an example of close cooperation with regulatory authorities to balance national security and AI safety.

The core driver of the security enhancement, GPT-Red, is an automated red-teaming model developed in-house by OpenAI. This model was used to systematically identify vulnerabilities that were difficult to find through traditional manual testing, training GPT-5.6 to have high resistance against prompt injection attacks. As a result, all models—Sol, Terra, and Luna—have formed a much more robust security layer than previous generations.

Cybersecurity Performance Benchmarks and Sol Model Capabilities

The top-tier model, GPT-5.6 Sol, recorded a score of 91.9% in Ultra mode on the cybersecurity benchmark 'Terminal-Bench 2.1'. This figure significantly exceeds the 82.0% of the previous model, GPT-5.5. According to an evaluation by the AI security research lab Irregular, the Sol model demonstrated excellent capabilities in performing complex hacking tasks and discovering unknown zero-day vulnerabilities in widely used software and mobile devices.

• GPT-5.6 Sol: Flagship model specialized in coding and cybersecurity ($30 per 1 million tokens)
• GPT-5.6 Terra: General-purpose model balancing performance and cost
• GPT-5.6 Luna: Efficiency-focused model optimized for high-speed processing ($1 per 1 million tokens)

According to the GPT-5.6 System Card released by OpenAI, these models were evaluated to possess 'High' level capabilities in cybersecurity and biological/chemical risk categories. However, as this falls below the 'Critical' threshold in OpenAI's own Preparedness Framework, commercial use was permitted subject to specific safeguards. External organizations, such as the UK AI Safety Institute (UK AISI), also participated in this process to verify the effectiveness of the safeguards.

The market is also paying attention to OpenAI's new pricing policy. GPT-5.6 introduced a model that sets cache-write costs at 1.25 times the input token rate, reflecting actual computing costs similar to Anthropic's approach. Meanwhile, skeptical views on the benchmark results exist in developer communities like Reddit, with a cautious stance prevailing that performance verification in real-world environments is necessary.

The Future of External Oversight and Adversarial Defense

OpenAI operates the 'Trusted Access for Cyber Security' program for security research. Through this, researchers in the biological and cybersecurity fields are granted access to the model's core functions to help identify potential risks in advance. This collaborative security model contributes to alleviating social anxiety caused by the rapid development of AI technology.

In conclusion, GPT-5.6 heralds the future of AI security by integrating automated red-teaming techniques into the development lifecycle. While prompt injection remains a challenge to be solved, the defense strategy through GPT-Red is evaluated as a significant advancement in defensive AI technology. OpenAI plans to sequentially roll out these models to ChatGPT users within the next few weeks.