OpenAI Discloses Cyber Pause, Teen ChatGPT Launch, and Wider Losses

In a blog post published Aug. 19, OpenAI said it paused some reinforcement learning workloads for two weeks after researchers determined that Astra, an unreleased algorithm more capable than GPT-5.6 Sol, qualifies as a critical cybersecurity risk under its Preparedness Framework. The framework defines a critical risk as a model that can find and exploit zero-day vulnerabilities in hardened systems without human help. OpenAI said the largest planned frontier reinforcement learning run remains on hold. The company also deployed activation classifiers, algorithms that regularly review its models' internal thought processes and tool interactions for signs of malicious activity. When an anomaly is found, a second, more advanced set of classifiers alerts researchers. OpenAI aims to generate alerts for suspicious behavior within 30 minutes and has instructed staff to pause such behavior within that time if they cannot rule out malicious intent. The monitoring overhead currently equals about 20% of the infrastructure allocated to the inference workloads being monitored, which could require price increases in the long term. OpenAI also narrowed some system access permissions, removed internal applications, and improved guardrails isolating high-risk workloads from the web.

Separately, OpenAI announced ChatGPT for Teens, a version of its chatbot for users aged 13 to 17. The company said it built guardrails to prevent conversations about violence, self-harm, sex, and mental disorders. If a child attempts to engage on a restricted topic and parental controls are enabled, the bot can alert parents. OpenAI does not use an age-verification tool; instead, it analyzes user inputs to detect age and will automatically switch on teen mode for younger users. The teen version includes a Study Mode for quizzes and visualizations, and a Study Hours feature that lets parents know how long their child is studying. In a press release, OpenAI said, "Teens should be able to use AI to learn, create, and explore. But that access should come with protections that reflect their developmental stage, reinforce real-world relationships, and support healthy use over time." The company said the design aims to support real-world relationships and healthy habits, with break reminders and product cues that identify ChatGPT as AI. The launch comes amid lawsuits alleging OpenAI's chatbot gave unsafe advice.

These announcements came as OpenAI's financial performance shows signs of slowing. According to a Wall Street Journal report citing undisclosed sources familiar with the company's finances, OpenAI generated $6.7 billion in sales in the quarter ending June 30, up from $5.7 billion in the first quarter, an 18% sequential increase. However, its operating loss climbed to $12.3 billion from $9.3 billion. The numbers are likely to disappoint investors ahead of a potentially imminent initial public offering, the Journal reported. The report came one day after Anthropic revealed that its revenue jumped more than 50% sequentially to $11.6 billion in the second quarter, with a small operating profit of $559 million. Anthropic said it achieved the profit through more efficient use of computing resources, though the company did not disclose its calculation method.

OpenAI has raised about $180 billion in funding and spent heavily on AI data centers and cloud contracts. Those deals, which have helped boost technology stocks, depend on OpenAI's ability to generate hundreds of billions of dollars in annual revenue. If it fails to meet growth targets and contractual obligations, share prices of companies such as Nvidia and Oracle could fall, the Journal reported. OpenAI also faces leadership turnover: it recently let go of Chief Revenue Officer Denise Dresser, following the departures of former Chief Operating Officer Brad Lightcap and Fidji Simo. President Greg Brockman, one of three original co-founders still involved, has become more involved in product and business development. OpenAI told investors privately that its growth accelerated after launching a new generation of models in July, but it did not provide numbers. It also released a "super app" integrating ChatGPT with its coding tool Codex and an AI-native web browser, and said the product is growing fast. OpenAI subsidizes hundreds of millions of non-paying users and recently lowered prices on two new models to attract corporate customers.