1. OpenAI Pauses Internal Astra Activities That Fall Short of New Safeguards as “Critical” Cyber Capabilities Cannot Be Ruled Out OpenAI has slowed development of Astra, an unreleased model designed for more advanced agentic coding and cybersecurity work, after preliminary evaluations raised the possibility that it had reached

2. Anthropic Adjusts Fable 5’s Biology Safety Classifier, Says Related Fallbacks Fell by About 85% Anthropic has refined the biology safety classifier for Claude Fable 5, aiming to reduce how often legitimate questions are mistakenly treated as risky.

3. Oracle Bans AI-Generated Code Submissions to OpenJDK, While LLMs May Still Be Used Privately for Debugging and Code Review Oracle has banned contributors from submitting AI-generated code or other AI-generated material to OpenJDK, the open-source Java project it stewards.

In Brief

• ByteDance Trains Model With Up to 10 Trillion Parameters ByteDance is in the early stages of pre-training a model that could reach 10 trillion parameters, according to three people familiar with the project. Its final size and release depend on how training progresses.

• Cloudflare Launches Cloud-Hosted Browser for AI Agents Cloudflare released Kitesurf, a browser for agents that can navigate websites and fill out forms without Chromium. It is free in beta through Browser Run, and Cloudflare claims it uses less CPU and memory for common agent tasks.

• Rippling Introduces Console for Tracking Enterprise AI Spending HR software company Rippling launched AI Spend Console, which links model usage and costs to employee or team output and includes a gateway for routing requests among models. Rippling says similar internal controls cut its token spending from 40% to about 15% of its R&D compensation budget without reducing token usage.

• SpaceX Shares Fall After AI Capital Spending Reaches $16 Billion SpaceX reported $7.8 billion in quarterly revenue and a roughly $541 million net loss, both better than analyst expectations, but its shares fell 10% after AI capital expenditure nearly doubled to $16 billion. Elon Musk said the company plans to expand computing capacity from 2 gigawatts at the end of 2026 to nearer 10 gigawatts than 5 by the end of 2027.

• Airbnb Begins Testing Optional Natural-Language Search Airbnb will test an AI search mode that returns visual results and personalized, AI-generated listing highlights while retaining its conventional search interface as an option. The company also says its support agent resolves nearly 45% of the cases it handles without human intervention, helping reduce support cost per booking by 16% year over year.

• New Mexico Court Raises Meta’s Child-Safety Penalties to $942 Million A New Mexico judge ordered Meta to pay another $567 million over alleged social-media harms, bringing total penalties in the case to $942 million. The order also restricts Like counts, overnight notifications, and monthly usage for minors in the state; Meta says it will appeal.

• OpenAI Partners With Psychological Association on Youth AI Safety OpenAI announced a partnership with the American Psychological Association to incorporate psychological research into responsible AI development and use among young people. The move follows lawsuits alleging that ChatGPT harmed users experiencing mental-health crises, allegations that remain subject to litigation.

• Permission Game Finds Players Missed One-Third of Agent Threats Across more than 40,000 runs of a browser game simulating AI-agent command approvals, players detected an average of 66.3% of threats. Commands hiding malicious behavior behind familiar script names were missed substantially more often, though the author cautioned that the game used artificial time pressure and an unusually high threat rate.

• Roku Adds a 24/7 Channel for AI-Generated Programming Roku launched Fairground AI Creator TV, a continuous free, ad-supported channel carrying videos from AI entertainment startup Fairground. Fairground says it pays participating creators and shares platform revenue, but does not identify the generation models or training-data rights behind the programming.

• OSReward Benchmark Finds AI Judges Too Lenient With Computer Agents Researchers introduced OSReward to test vision-language models that judge whether computer-using agents completed tasks, finding a systematic tendency to classify failed runs as successes. They also released a 100,000-example dataset and two open reward models that they report can match commercial judges at lower cost.

Read the full edition →