The Battle Between Local and Cloud Computing

Η εταιρεία ενισχύει τις δικλίδες προστασίας μετά το περιστατικό με το Hugging Face και ετοιμάζεται να διαθέσει το νέο μοντέλο τεχνητής νοημοσύνης σε επιλεγμένους δοκιμαστές.

Artificial intelligence recently breached a crucial internal OpenAI boundary. Basic safeguards against malicious requests simply no longer suffice. The company officially recognized Astra as its first model possessing critical cybersecurity capabilities. During rigorous testing, the system autonomously identified previously unknown vulnerabilities. It successfully constructed functional exploits. Furthermore, it expertly combined multiple errors into comprehensive exploitation chains against secured systems.

OpenAI has announced the release of its Astra AI model, which is the first to meet the company's threshold for 'critical' cyber capabilities. The model can independently identify and exploit previously unknown software vulnerabilities, prompting OpenAI to implement safety measures and pause training workloads temporarily. Astra's advanced capabilities include chaining multiple exploits to penetrate systems, outperforming industry leaders like GPT-5.6 Sol and Anthropic's Mythos on cybersecurity benchmarks. While OpenAI emphasizes safeguards such as a 'misalignment monitor' to prevent misuse, cybersecurity experts caution that AI's growing hacking abilities pose significant risks to unprepared systems. The company is offering early access to select partners through its Daybreak Blue program to strengthen defenses before broader release. This development highlights ongoing industry challenges in balancing AI innovation with security protocols, as similar incidents have been reported by competitors like Anthropic and Meta.

Το νέο AI μοντέλο ξεπέρασε το όριο ασφαλείας της εταιρείας αναπτύσσοντας τη δυνατότητα να ανακαλύπτει και να αξιοποιεί zero-day exploits.

OpenAIのAIモデルAstraが10の数学難問を解決。