OpenAI Rolls Out GPT‑6 Astra, Claims Full Score on ExploitBench Amid New Security Restrictions

OpenAI announced Thursday that its latest language model, GPT‑6 Astra, has achieved a perfect rating on the ExploitBench benchmark, a test suite used to gauge a system's susceptibility to known cybersecurity exploits. The company described Astra as the "world's most intelligent and aligned model," emphasizing its advanced safety controls and its recent classification as having reached the "Critical" threshold for cybersecurity capability.

The ExploitBench result, a 100% success rate in avoiding exploit triggers, marks a notable milestone for large‑scale AI, which has traditionally struggled to balance raw capability with robust security safeguards. Industry observers note that the benchmark assesses whether a model can be coaxed into generating malicious code or facilitating attacks, making the score a key indicator of alignment progress.

OpenAI's disclosure follows a policy shift announced earlier this month, in which the firm said it would no longer entertain proof‑of‑concept (PoC) exploit requests from external researchers. The move, described by OpenAI as a protective measure, aims to limit the distribution of detailed attack vectors that could be misused if released publicly. Critics argue that restricting such collaboration could slow the discovery of vulnerabilities, while the company maintains that the decision reflects a responsible approach to managing emerging risks.

GPT‑6 Astra builds on the architecture of its predecessor, GPT‑5, incorporating larger training datasets and refined alignment techniques such as reinforcement learning from human feedback (RLHF) and adversarial training. The "Critical" cybersecurity capability label suggests that the model meets internal thresholds for detecting and mitigating high‑severity threats, a standard OpenAI introduced after earlier models exhibited occasional lapses in safety.

Experts in AI safety see the announcement as a test case for how developers can embed stronger defensive layers without compromising performance. "Achieving a perfect score on a benchmark like ExploitBench signals that alignment research is maturing," said a cybersecurity analyst who follows AI developments, though the analyst cautioned that benchmarks are only one measure of real‑world resilience.

The broader AI community is watching OpenAI's stance on PoC requests closely, as collaborative vulnerability research has historically helped harden software and models. By limiting external probing, OpenAI may be prioritizing immediate risk mitigation over longer‑term transparency. The company has not indicated whether it will open a private bug‑bounty program or other controlled channels for security researchers.

OpenAI's release of GPT‑6 Astra, coupled with its new security policy, underscores the growing tension between rapid AI advancement and the need for robust safeguards. As AI systems become more integral to business, government, and everyday applications, the balance between openness and protection will likely shape regulatory discussions and industry standards in the months ahead.