OpenAI’s GPT-6 Astra and Google’s Gemini 3.8 Flash Cyber have demonstrated the ability to autonomously develop software exploits. GPT-6 Astra achieved a 100% score on OpenAI’s ExploitBench. This result marks a significant leap from the 78.5% score recorded by its predecessor.
Google is reportedly positioning its Gemini 3.8 model specifically for vetted cybersecurity defenders. This move highlights the dual-use dilemma as offensive AI capabilities outpace current safety protocols. Existing safeguards rely on access gates and classifiers rather than hard capability limits to control dangerous functionalities.
The advancement signals the arrival of AI-generated cyberweapons in the information technology sector. While these tools assist in vulnerability discovery, they also establish a new frontier for offensive operations. This shift carries significant implications for the entire security industry.