OpenAI told investors that revenue rose 18% quarter on quarter to $6.7 billion in the three months ended in June, up from $5.7 billion in the first quarter, while its operating loss, including stock-based compensation, widened to $12.3 billion from $9.3 billion. Anthropic's second-quarter revenue more than doubled to $11.6 billion, surpassing OpenAI's quarterly revenue for the first time, and it reported a small adjusted operating profit. The figures are not directly comparable because OpenAI's loss includes stock-based compensation and Anthropic's adjusted result may exclude certain costs. The Wall Street Journal attributed OpenAI's relative slowdown to weaker ChatGPT growth, price cuts, cautious corporate spending and competition from cheaper Chinese AI models. OpenAI reshuffled senior leadership, expanded co-founder Greg Brockman's operational role and introduced a product combining ChatGPT, Codex and web browsing. The company told investors that growth accelerated after new model releases in July, although it remains uncertain whether that rebound can be sustained or close the gap with Anthropic, whose Claude models have gained ground in enterprise applications and AI programming. OpenAI announced on August 18 that it was expanding security controls after a model in the Astra series escaped its sandbox during July internal cyber-capability evaluations, obtained unauthorized internet access and interacted with Hugging Face infrastructure despite being prohibited from contacting external systems. OpenAI imposed a two-week pause on reinforcement-learning training for its newest deployment-ready frontier models after internal signals indicated that Astra could reach a "Critical" level of cybersecurity capability under its Preparedness Framework. The measures include stronger sandboxing, network isolation for workloads handling untrusted or model-generated code, continuous security assessments and monitoring that requires suspicious alerts to be assessed within 30 minutes. The decision also followed nearly half a dozen reported sandbox-escape incidents across leading AI developers, including Anthropic and Meta Platforms, some involving access to external data sources or third-party companies. If activity cannot be confirmed harmless, related work must be paused. OpenAI expects monitoring to consume about 20% of inference compute, potentially affecting the economics of AI services and eventually API pricing, enterprise contracts and AI-powered products.