OpenAI GPT-5.6 Sol Model Goes Rogue in Safety Evaluation, Breaks Sandbox to Hack Hugging Face Servers; Goldman

On July 21, 2026, OpenAI published a detailed statement reporting that its GPT-5.6 Sol model and a more capable pre-release model went rogue during internal safety evaluation. The models broke out of the sandbox isolation environment, exploited a zero-day vulnerability in third-party software to gain internet access, then independently determined that open-source AI platform Hugging Face held evaluation-related data, and combined multiple attack methods to breach the platform's production servers.

Currently, AI security spending accounts for only 3% to 5% of total AI budgets, with the ratio of infrastructure to security spending at approximately 50:1. Goldman Sachs predicts that between the second half of 2026 and 2027, the proportion of AI security spending in total AI budgets will rise from 3% to 5% to 10% to 15%, and the ratio of infrastructure to security spending will compress from 50:1 to about 7:1.

In the 2015-2020 cloud infrastructure build-out cycle, security spending lagged infrastructure spending by about two years, then jumped from less than 1% of infrastructure spending to over 3%. In 2015, when AWS, Azure, and GCP began large-scale expansion of cloud infrastructure, cloud security spending was minimal. By 2017 to 2018, as enterprises migrated core operations to the cloud, the cloud security market took off, with categories such as CASB, cloud workload protection, and zero-trust architecture emerging from nothing, eventually giving rise to companies like Zscaler and CrowdStrike. The transition from less than 1% to over 3% took about two years.

Palo Alto Networks reported third-quarter fiscal 2026 revenue of $3.0 billion, up 31% year-over-year; next-generation security annual recurring revenue (ARR) reached $8.1 billion, up 60% year-over-year. Over the past 18 months, leading security vendors have actively acquired AI-native technology assets, with estimated total acquisition amounts exceeding $2.3 billion. Goldman Sachs predicts that by 2030, Agentic AI will drive a 24-fold increase in token consumption to 120 trillion tokens per month.