{
  "id": 4926477,
  "title": "OpenAI Is About to Release Its First AI Model With ‘Critical’ Cyber Abilities",
  "url": "https://urgent.news/2026/09/01/openai-is-about-to-release-its-first-ai-model-with-critical-cyber",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-01T20:00:00.000Z",
  "source": {
    "name": "Wired",
    "slug": "wired",
    "url": "https://www.wired.com/story/openai-astra-first-ai-model-with-critical-cyber-abilities/"
  },
  "original_language": "en",
  "account": "OpenAI is set to unveil its first AI model, Astra, that possesses critical cybersecurity capabilities. The company has determined that Astra meets the thresholds and protocols established in its preparedness framework, which signals a new level of risk. To reach this critical cyber threshold, Astra can independently identify and exploit previously unknown vulnerabilities in real-world software. Following protocol, OpenAI halted further development of Astra and a future AI model for several weeks until appropriate safeguards and security measures could be implemented. The multi-week pause proved productive, and OpenAI is now confident in releasing Astra safely. This announcement comes at a time when Silicon Valley is grappling with the advanced cybersecurity abilities of cutting-edge AI models, while striving to assure users, lawmakers, and other companies that they can be controlled. Incidents involving AI models exploiting vulnerabilities in testing environments have raised concerns across the industry. To mitigate potential misuse, OpenAI has implemented a multi-step approach, including a new \"misalignment monitor.\" This monitor is designed to refuse to answer if Astra is asked to find an exploit in a real-world software system. OpenAI has also made Astra more resistant to jailbreaking attempts and successfully refused unsafe queries in tests at a significantly higher rate compared to previous models. However, there is a possibility that the misalignment monitor might occasionally flag legitimate activities as potential cyber misuse, causing it to slow down, pause, or stop. Digital infrastructure providers, including Cisco, Cloudflare, and Palo Alto Networks, will have early access to a less restricted version of Astra with enhanced cyber capabilities through OpenAI's Daybreak program. The aim is to ensure these companies can utilize advanced AI models like Astra to strengthen their defenses before models of similar capabilities are made widely available. OpenAI's Astra outperforms industry-leading AI models such as GPT-5.6 Sol and Anthropic's Mythos on cybersecurity benchmarks, scoring 100 percent on ExploitBench.",
  "summary": "The company will give select partners early access to its Astra AI model—so they have time to shore up their defenses.",
  "key_points": [
    "OpenAI to release Astra, first AI model with critical cybersecurity abilities",
    "Astra can independently identify and exploit unknown vulnerabilities in real-world software",
    "OpenAI implemented safeguards, including a misalignment monitor, before Astra's release"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 3,
    "also_reported_by": [
      {
        "outlet": "Wired Business",
        "title": "OpenAI Is About to Release Its First AI Model With ‘Critical’ Cyber Abilities",
        "url": "https://urgent.news/2026/09/01/openai-is-about-to-release-its-first-ai-model-with-critical-cyber-4932328",
        "published": "2026-09-01T20:00:00.000Z"
      },
      {
        "outlet": "Techmeme",
        "title": "OpenAI says Astra is its first model to reach its \"Critical\" cyber threshold and warns safeguards may mistakenly flag legitimate activity as cyber misuse (Ina Fried/Axios)",
        "url": "https://urgent.news/2026/09/01/openai-says-astra-is-its-first-model-to-reach-its-critical-cyber",
        "published": "2026-09-01T20:15:01.000Z"
      }
    ]
  },
  "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."
}