OpenAI's Astra Surpasses Cybersecurity Benchmarks Amid Industry Updates

OpenAI's latest language model, Astra, has achieved a significant milestone by surpassing the 'Critical' cybersecurity threshold in the company's Preparedness Framework. The model, which autonomously found and exploited two zero-day vulnerabilities in modified tests, is set to be released soon, albeit with restrictions on its most advanced capabilities.

OpenAI's Astra Hits 'Critical' Cyber Tier

OpenAI announced that Astra is the first large language model (LLM) to score a perfect ExploitBench, a benchmark designed to evaluate a system's ability to find and exploit security vulnerabilities. This achievement highlights the growing sophistication of AI in cybersecurity, raising both opportunities and concerns about the potential for misuse.

Trump DOJ Supports OpenAI in Copyright Case

In a related development, the Trump administration has filed a statement of interest in a Manhattan federal court, backing OpenAI's fair-use defense against the New York Times. The brief argues that training AI models on publicly available internet material is a fair use, marking the first time the administration has intervened in the wave of copyright lawsuits against AI labs.

ChatGPT Integrated into Epic EHR System

OpenAI also launched ChatGPT Health, integrating the chatbot into Epic's electronic health record (EHR) system. This read-only integration allows clinicians to summarize appointment notes, lab results, medications, and specialist documents, streamlining their workflow without the ability to write back to the system. A new Healthcare Public Data plug-in further enhances the tool's utility.

Google Set to Unveil Gemini 3.8 Flash

Meanwhile, Google DeepMind is preparing to unveil Gemini 3.8 Flash, codenamed 'skimaki,' on September 2. The model, which has been tested on the internal Jetski coding platform, aims to compete with Anthropic's Claude Fable 5. According to reports, Gemini 4 is showing promise in pre-training evaluations but still requires additional post-training work.

Fake Sites Generate 215K AI-Only Pages

Researcher Trellner uncovered a concerning trend in AI-generated content, finding that three domains—wifitalents.com, worldmetrics.org, and gitnux.org—collectively host 215,128 machine-generated buying guides and other pages. These sites, which are not among the top 100,000 globally, were cited in nearly 60% of the 7,534 citations generated from 380 software queries submitted to Perplexity's sonar and sonar-pro.

References

• AI News Today, September 2: Top Stories | AI Weekly