OpenAI First Tests GPT-5.6 Sol With Restricted Access and Strong Cyber Defenses
OpenAI on Friday released three versions of the GPT-5.6it was called Sol, Terra, and Lunasuch as a limited preview of a small number of companies as part of ongoing engagement with the US government.
While the Sol is the latest flagship model and is more powerful, the Terra strikes a balance between efficiency and power, and the Luna is well-tuned for speed and affordability.
“GPT-5.6 Sol debuts with our strongest security stack to date. We’ve strengthened defenses against high-risk workloads, critical Internet applications, and repeated exploits, and spent weeks finding vulnerabilities, testing our system, and hardening it against real-world attacks,” OpenAI said.
The model has also been called “the most efficient model yet” for cybersecurity, making it particularly suitable for vulnerability and exploit research. In ExploitBench , GPT-5.6 Sol competes with Anthropic Mythos Preview using only one-third of the output tokens, noted OpenAI.
The goal, he added, is to enable access to legitimate work such as code reviews, vulnerability research, patch development, debugging, security education, and defense testing, while enforcing strict security rules that block offensive activity and quickly fix newly discovered jailbreaks. This includes controversial efforts to break the model and reject what it describes as “prohibited online help.”
“As these capabilities continue to evolve, our priority is to ensure that they reach and benefit defenders, who can use these tools to find vulnerabilities, develop patches, and strengthen systems more broadly,” explained the artificial intelligence (AI) company.
That said, OpenAI also warns that there may be situations during the preview phase where users may encounter safeguards that block or reject legitimate requests, or have their requests temporarily suspended for further review, due to the “dual-use” nature of the technology.
According to OpenAI’s GPT-5.6 Preview System card, although the model is very adept at detecting code vulnerabilities and developing exploits, the capabilities do not extend to conducting independent, persistent attacks against strong targets or equipping those cyber vulnerabilities in actual attacks.
“Separate tests examined the wrong behavior in the agent’s coding activities and found that GPT-5.6 shows a greater tendency than GPT-5.5 to exceed the user’s intention, including taking or attempting actions that the user did not request, although the absolute values remain low,” it pointed out.
Testing GPT-5.6 Sol against widely used software projects hardened using VulnLMP, an internal OpenAI framework designed to test exploit chain development against real-world targets, found the model to generate reliable memory security traces, some of which could lead to the disclosure, modification, or control of flow corruption.
“This suggests that large parts of real-world risk research are becoming increasingly automated when models are coupled with the use of tools, building systems, and validation infrastructure,” the tech upstart said.
OpenAI aims to make GPT-5.6 Sol, Terra, and Luna generally available in the coming weeks, and preview the model’s capabilities to the US government. It also presents a limited preview to a small group of trusted partners whose participation is approved by the government before a wider launch.
Earlier this month, US President Donald Trump signed an executive order on AI and cybersecurity, calling for the creation of a framework that gives the federal government the power to test the capabilities of AI models and determine which ones qualify as “covert frontier models,” a term for AI systems with advanced cyber capabilities.
The surprising release comes days after the company released an improved version of its GPT-5.5‑Cyber model to trusted defenders as part of the Daybreak program and launched a new project called Patch the Planet in collaboration with Trail of Bits to help protect open source projects.
It also follows the US government’s decision to allow Anthropic to release its Mythos AI model to a group of about 100 trusted companies and federal government agencies that “operate and protect critical infrastructure,” more than two weeks after powerful models focused on cybersecurity were released to the market.
“We are restoring access to these organizations immediately, and we continue to work with the government to expand access to Mythos 5 and make Fable 5 available for use again,” Anthropic said in a statement sent to X.
