Google launches Playground: prompt your own games, no code
Google turned game-making into a chat window today, while Anthropic opened the most permissive tier of its cyber program to government-vetted defenders and put its AI watermark detector in front of everyone.
Google launched Playground, a browser platform where anyone describes a game in text prompts and plays it — no coding, live today in the US for users 18 and over. Creation works from a blank canvas or starter prompts: you type what you want, then tweak physics, rewrite rules, or swap characters in the same conversation until it plays the way you imagined. Finished games stay private, ship as a shareable link, or go public in the Explore gallery, with leaderboards and multiplayer in select genres, Play Games profiles for handles and follows, and safety screening on everything published. A Google spokesperson told The Verge the platform runs on Gemini, Nano Banana, and Lyria behind a custom game-making harness; it's free, with Google One subscribers getting higher weekly token limits by plan tier. The more interesting piece is the ladder behind it: Unity Spark, an integration that brings the Unity runtime's professional 3D mechanics into Playground, enters closed beta later this year, so a prompt-made prototype can grow into a real game without leaving the platform. Google is arriving late to a race Meta Horizon Create Studio and Roblox are already running, but browser-based distribution means Playground games open on any phone or laptop — no install, no engine, and Google's model stack doing the heavy lifting.
Anthropic's Cyber Verification Program now hands vetted defenders its strongest Claude models in three tiers — and the top tier is co-vetted with the US government. Defense Access is open to corporate security teams, agencies, universities, critical infrastructure operators, and individual researchers; Red Team Access allows authorized attack simulations for organizations only; Specialized Access covers flight controls, power grids, and banking infrastructure, with every applicant screened alongside the government. The models include Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 with reduced blocking classifiers, and Anthropic says mandatory data retention applies for misuse monitoring. The numbers justify the loosening, if the reporting holds: predecessors in Project Glasswing found at least 129,000 confirmed vulnerabilities between April and July 2026, over 33,000 of them high or critical, and Anthropic claims true impact runs at least five times the survey figures. On its CyScenarioBench, models with no program access were blocked at the first prompt in all 50 trials, while Red Team tier access completed 34 of 50. We covered the safeguards shift that led here — Anthropic ships Sonnet 5.5 as cyber safeguards move down a tier.
Google opened its upgraded SynthID detector to everyone, everywhere, in English — you can now check any image, video, or audio file for AI watermarks at synthid.com. Google says 180 billion images and videos and 240,000 years of audio already carry SynthID marks, with over a million verification requests a day flowing through Search, Gemini, and Chrome. The upgrade matters because the checker no longer only recognizes Google's watermarks: OpenAI, NVIDIA, and Kakao are in as partners, with Apple expected soon, so one site can now probe media from several generators at once. The limits are the other half of the story — the verdict is a yes/no with no highlighted region, logins are required at roughly ten checks a day, and content from non-partners like Meta or open-weight models passes undetected. Detection is going consumer; provenance is still a consortium.
What to watch: whether Unity Spark's closed beta keeps its "later this year" date, and how fast Playground expands past US borders.
Is a government co-signed tier the right line for handing out unhobbled frontier models — or does it just move the accountability next door? Tell us in the comments.
