{
  "id": 5538500,
  "title": "OpenAI launches GPT-6 Astra, its first model to cross a critical cybersecurity threshold",
  "url": "https://urgent.news/2026/09/04/openai-launches-gpt-6-astra-its-first-model-to-cross-a-critical",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-04T09:46:11.000Z",
  "source": {
    "name": "Computerworld",
    "slug": "computerworld",
    "url": "https://www.computerworld.com/article/4218691/openai-launches-gpt-6-astra-its-first-model-to-cross-a-critical-cybersecurity-threshold-3.html"
  },
  "original_language": "en",
  "account": "OpenAI introduced GPT-6 Astra on Thursday, revealing it has surpassed the \"Critical\" cybersecurity risk level under its Preparedness Framework, a classification requiring additional deployment restrictions. The new model will initially be available to a select group of organizations before becoming accessible to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS. Enterprise administrators must manually enable Astra in their workspace, as access is initially disabled upon launch.\n\nDevelopers can access Astra via the API or Amazon Bedrock, priced at $10 per million input tokens and $50 per million output tokens. Astra Pro, a variant for Pro, Business, and Enterprise users, is also available. The company claims Astra has achieved a perfect score of 100% on the ExploitBench benchmark, up from 78.5% for its predecessor, GPT-5.6 Sol. Astra also scored 42.4% success against 30.3% for Sol on ExploitGym, a broader exploit-development benchmark, with fewer output tokens utilized.\n\nWhile Astra's ability to identify and develop zero-day exploits can aid defenders in finding and patching weaknesses, it also necessitates stronger safeguards. OpenAI tested Astra on vulnerabilities disclosed in the three months prior to its launch, finding two new zero-day vulnerabilities during this test. They have since disclosed both to the respective software creators.\n\nSanchit Vir Gogia, chief analyst at Greyhound Research, explained that the Critical label is a disclosure event rather than a capability event. Astra's capability remained unchanged between August and September, but testing changed, making it the only frontier model whose cyber capability is explicitly measured. This measurement reveals that unlabelled models behind enterprise credentials have never been evaluated, making them potentially less safe.\n\nOpenAI plans to tighten restrictions on advanced offensive tasks for the public version of Astra, such as generating proof-of-concept exploits. However, they intend to ease these limitations for vetted defenders through OpenAI Daybreak in the coming weeks. The launch comes after Anthropic briefly pulled its Fable and Mythos models from export markets due to similar concerns.\n\nThe shift in governance is from the model to the harness surrounding it, as reasoning now leads to state changes instead of information problems. A wrong chatbot answer is now an operational event rather than an information issue. This change means the relevant question is no longer which model is approved, but how much damage a specific identity can cause before a control intervenes.",
  "summary": "OpenAI launched GPT-6 Astra on Thursday, disclosing that the new flagship model has crossed the “Critical” threshold for cybersecurity risk under its Preparedness Framework, a classification the company said triggers additional deployment restrictions. “GPT‑6 Astra is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro,…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}