OpenAI on Friday launched three variations of GPT-5.6, referred to as Sol, Terra, and Luna, as a restricted preview to a small variety of firms as a part of an ongoing engagement with the U.S. authorities.
Whereas Sol is the newest flagship mannequin and essentially the most highly effective, Terra strikes a steadiness between effectivity and energy, and Luna is fine-tuned for pace and affordability.
“GPT‑5.6 Sol launches with our most sturdy security stack up to now. We strengthened protections for higher-risk exercise, delicate cyber requests, and repeated misuse, and spent a number of weeks discovering weaknesses, pressure-testing our system, and hardening it in opposition to real-world assaults,” OpenAI mentioned.
The mannequin has additionally been touted because the “most succesful mannequin but” for cybersecurity, making it way more appropriate for vulnerability analysis and exploitation. On ExploitBench , GPT‑5.6 Sol is aggressive with Anthropic Mythos Preview utilizing solely about one-third of the output tokens, OpenAI famous.
The objective, it added, is to allow entry to reliable work reminiscent of code assessment, vulnerability analysis, patch growth, debugging, safety schooling, and defensive testing, whereas imposing sturdy guardrails that block offensive exercise and swiftly remediating newly found jailbreaks. This consists of adversarial makes an attempt to jailbreak the mannequin and refuse what it describes as “prohibited cyber help.”
“As these capabilities proceed to advance, our precedence is to ensure they attain and profit defenders, who can use these instruments to seek out weaknesses, develop patches, and strengthen programs extra broadly,” the bogus intelligence (AI) firm defined.
That mentioned, OpenAI can also be warning that there could also be situations throughout the preview part the place customers could encounter safeguards that block or refuse reliable requests, or have their requests paused for added assessment, owing to the “dual-use” nature of the know-how.
In line with OpenAI’s GPT-5.6 Preview System Card, though the mannequin is more proficient at discovering vulnerabilities in code and growing exploits, the capabilities don’t lengthen to finishing up autonomous, end-to-end assaults in opposition to hardened targets or weaponizing these cyber vulnerabilities in actual assaults.
“Separate evaluations examined misaligned habits in agentic coding duties and located GPT-5.6 exhibits a higher tendency than GPT-5.5 to transcend the person’s intent, together with by taking or trying actions that the person had not requested for, although absolute charges stay low,” it identified.
An analysis of GPT-5.6 Sol in opposition to extensively deployed hardened software program tasks utilizing VulnLMP, which is OpenAI’s inside framework designed to check end-to-end exploit chain growth in opposition to real-world targets, has discovered the mannequin to provide credible reminiscence security leads, a few of which might result in disclosure, mutation, or management circulate corruption.
“This implies that substantial elements of actual world vulnerability analysis have gotten more and more automatable when fashions are paired with software use, construct programs, and verification infrastructure,” the tech upstart mentioned.
OpenAI intends to make GPT‑5.6 Sol, Terra, and Luna typically out there within the coming weeks, and it previewed the mannequin capabilities to the U.S. authorities. It is also launching a restricted preview for a small group of trusted companions whose participation has been authorized by the federal government earlier than a broader launch.
Earlier this month, U.S. President Donald Trump signed an government order on AI and cybersecurity, calling for the creation of a framework that grants the federal authorities the power to guage AI fashions’ capabilities and decide which qualify as “lined frontier fashions,” a designation for AI programs with superior cyber capabilities.
The staggered launch comes days after the corporate launched an improved model of its GPT‑5.5‑Cyber mannequin to trusted defenders as a part of the Dawn initiative and launched a brand new undertaking referred to as Patch the Planet in collaboration with Path of Bits to assist safe open-source tasks.
It additionally follows the U.S. authorities’s determination to allow Anthropic to launch its Mythos AI mannequin to a gaggle of about 100 trusted firms and federal authorities companies that “function and defend vital infrastructure,” greater than two weeks after the highly effective cybersecurity-focused fashions had been pulled from the market.
“We’re restoring entry for these organizations shortly, and we’re persevering with to work with the federal government to broaden entry to Mythos 5 and make Fable 5 out there for basic use once more,” Anthropic mentioned in an announcement posted on X.
