OpenAI Unveils GPT-6 Astra, Achieves 100% Exploit Score

"Astra is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work," OpenAI said — and by several measures the company has the test scores to back that claim.

GPT‑6 Astra’s headline numbers: FrontierMath, ARC‑AGI‑3 and ExploitBench

OpenAI on Thursday unveiled GPT‑6 Astra and published a string of benchmark results that position the model as a new frontier for professional and cyber capabilities. OpenAI said Astra "saturates FrontierMath Tier 4 with a 98% score," "saturates ARC-AGI-3 with a 99.9% score," and achieved a perfect 100% on ExploitBench. By comparison, OpenAI reported that its previous frontier cyber-capable model, GPT‑5.6 Sol, scored 78.5% on ExploitBench.

ExploitBench and Astra’s exploit-development performance

ExploitBench, OpenAI notes, evaluates a model’s ability to turn known software vulnerabilities into working exploits. In testing covering flaws from July and August 2026, Astra reportedly delivered "substantially higher arbitrary code-execution rates" than GPT‑5.6 Sol, including behavior that leveraged two zero-day vulnerabilities in unspecified software. OpenAI further said Astra can, if allowed to run without safeguards, use previously unknown vulnerabilities to achieve code execution in hardened browsers and to develop privilege‑escalation exploits for hardened operating systems.

Release constraints: secure review, PoC refusals, and safety checks

Recognizing the dual-use risk — where capabilities that accelerate defensive work can also make offensive actions easier — OpenAI described a constrained release for Astra. The company said the version being released is limited to "secure code review and patching" and that it will "refuse to comply with prompts related to creating proof-of-concept (PoC) exploits for vulnerabilities." OpenAI also listed additional safeguards: "stronger model robustness to better tackle jailbreaks," "more context to its monitoring systems," and "extra safeguards to help detect and contain misalignment."

OpenAI warned that these safety checks can at times interrupt legitimate defensive work; in those cases "the user will be prompted to review the action before continuing." The company added that Astra is "more likely" to operate within the confines set by the user and implied by its environment, and that running with additional security measures by default "yielded even stronger performance."

Daybreak for Frontline Defenders, $1 billion commitment, and MS‑ISAC pilot

OpenAI announced Daybreak for Frontline Defenders, a global project the company says "aims to commit $1 billion to help defenders use frontier AI cyber capabilities to safeguard essential services against cyber attacks." The initiative targets critical infrastructure sectors, explicitly naming water systems, electricity providers, state and local governments, banks, non‑profits, and open-source maintainers. OpenAI also announced a pilot with the U.S. Multi-State Information Sharing and Analysis Center (MS‑ISAC) to equip an initial group of public-sector and water-system defenders with Daybreak access, guided training, and hands-on assistance.

The company framed the program as time-sensitive: "We have a defender's window: a narrowing opportunity to use AI to close security gaps before attackers seize them," OpenAI said.

How technologists, public‑sector and water defenders, and open‑source maintainers are positioned

• Technologists and security teams: They will encounter a model that, per OpenAI, can accelerate secure code review and patching but that also has exploit‑level capability when unrestricted; they must account for safety check interruptions, because Astra will prompt users to review potentially sensitive actions.
• Public‑sector and water-system defenders: The MS‑ISAC pilot specifically targets this group with Daybreak access, training, and assistance — an immediate pathway to evaluate Astra under the constrained configuration OpenAI described.
• Open‑source maintainers and resource‑limited organizations: Daybreak’s stated $1 billion commitment is aimed in part at these groups, offering subsidized access, hands‑on training, and technical assistance intended to let defenders use frontier AI capabilities for protective work.

OpenAI said Astra is rolling out first to a small set of organizations and is expected to become available to all ChatGPT Plus, Pro, Business, and Enterprise users and via the OpenAI API, Microsoft Azure, and Amazon Web Services Bedrock. The company also signaled further loosening of restrictions in the near term: "Through OpenAI Daybreak, we plan to expand access and roll out less restrictive safeguards in the coming weeks," the company said, with the intention of enabling defensive workflows such as vulnerability and proof‑of‑concept validation, malware analysis, and detection engineering.

There is a clear tension at the center of OpenAI’s announcement: Astra’s benchmarked capability includes behavior that can produce working exploits and leverage zero‑day flaws, yet the released configuration is explicitly designed to deny exploit-generation prompts and to prioritize secure review workflows. OpenAI’s stated $1 billion Daybreak commitment and the MS‑ISAC pilot place frontline defenders at the front of that experiment; the coming weeks will show how successfully Astra’s safeguards permit defensive work without enabling offensive abuse.

https://thehackernews.com/2026/09/gpt-6-astra-scores-100-on-exploitbench.html