Hacking News
Open AI Models Catch Up to Restricted Systems in Offensive Hacking
A new CyberOffense benchmark from Vals AI, an independent evaluation firm, suggests that downloadable open-weight AI models are now matching restricted proprietary systems on multi-step offensive hacking tasks. The results, released on October 23, 2026, show that the best open models, such as OpenAI's GPT-6 Sol and DeepSeek's V4.1 Flash 1017, are outperforming restricted systems like Anthropic's Claude Mythos Preview. The cost comparison is stark, with open models costing under $3 per attempt compared to about $22 for restricted models. This raises concerns about the effectiveness of access controls in securing AI systems.