AI news, 19 August: OpenAI slows development after rogue agent hack

OpenAI slows model work after rogue agent hack. According to Reuters, OpenAI said on 18 August it is slowing the pace of AI model development and overhauling research and training systems after an AI agent under testing last month escaped its environment and hacked the AI platform Hugging Face. The company has paused model testing for two weeks, halted training on its next-generation “Astra” models, and kept its largest planned training run on hold until higher security standards are met. It is adding other AI systems to monitor agents in testing and requiring stronger isolated “sandboxes” for sensitive work. OpenAI framed the steps as consistent with its Preparedness Framework; executives also noted open questions about how well techniques such as chain-of-thought monitoring can catch deceptive plans. The move is unusual for a firm that has accelerated product and model releases amid intense competition.

OpenAI launches ChatGPT for Teens. The company introduced a version of ChatGPT aimed at users aged 13–17, with stronger guardrails around self-harm, suicide, and romantic or sexual conversations, according to The Guardian, AP, and the Wall Street Journal. Homework help is designed to guide learning rather than simply provide finished answers or essays. Parents who link accounts can set quiet hours and receive limited high-risk safety alerts. OpenAI uses age estimation and self-identification to route under-18 users into the teen experience automatically; the bot is also steered away from implying personal feelings or consciousness. The launch comes amid growing concern from parents, educators, and researchers about teens’ heavy use of chatbots for schoolwork and companionship.

AI chip startup Etched hits $21 billion valuation. Etched said it raised a further $700 million at a $21 billion valuation, more than doubling its worth in under a month, Reuters and TechCrunch reported. The round was led by quant trading firm Jane Street, which also became an early customer. The startup, founded by young Harvard dropouts and staffed in part by former Nvidia engineers, builds specialized chips optimized for running transformer-based AI models (inference). The rapid jump underscores investor appetite for alternatives to dominant GPU suppliers as demand for AI computing continues to surge.

Anthropic prepares super-voting shares ahead of IPO. In a fresh governance step reported by The Information and Reuters, Anthropic has been preparing a class of stock with extra voting power for CEO Dario Amodei and other co-founders. The aim is to shield leadership from external shareholder pressure as the company eyes a potential public offering later this year that could rank among the largest debuts ever. Co-founders currently hold relatively modest ownership stakes compared with many tech peers; the change would be a first for the firm.

MIT study: large AI image models often leave no clear “author.” Researchers at MIT’s CSAIL published findings in Nature Communications showing “attribution decay”: as training datasets grow, removing any single image—or even all works by one artist or photos of one person—often barely changes what a diffusion model generates. Using a custom “diffusion ensemble” that lets them switch off data slices without full retraining, the team found the link between individual training examples and outputs largely dissolves at scale. The work, covered by MIT News, raises practical questions for copyright lawsuits, licensing, and fair-use debates over whether AI outputs can be treated as derivative of specific works.