Google opens Genie 4 on the Gemini API for 720p, 24fps playable worlds Google released Genie 4 on the Gemini API, letting developers generate and steer 720p, 24fps playable worlds from their own code; a day earlier the model was available only inside Project Genie for eligible Google AI Ultra subscribers. Google pitches the API access for prototyping and for training robotics and AI agents. A day earlier, Genie 4 was reachable only inside Project Genie for eligible Google AI Ultra subscribers. Developers can now generate and steer worlds from their own code, which Google pitches for prototyping and for training robotics and AI agents. Read: A day earlier, Genie 4 was reachable only inside Project Genie for eligible Google AI Ultra subscribers. Developers can now generate and steer worlds from their own code, which Google pitches for prototyping and for training robotics and AI agents. Read: One HTTP/2 bug affects every release since 1.9.5 in 2015, and a chained Linux kernel bug turns the foothold into root even inside a Docker container. Origins reachable only through Cloudflare's proxy are shielded from the HTTP/2 and HTTP/3 attacks. Read: The one-byte heap overflow is reachable without a login on default Rails and Sinatra setups, and Ruby also ships inside self-hosted products such as GitLab and GitHub Enterprise Server. Fixes are in Ruby 3.3.14, 3.4.13 and 4.0.10, with the exploit due in a week. Read: More than 24,000 businesses had given agents read access to Treasury data. Agents can now pay vendors, convert currency and transfer funds themselves, with refunds, payouts and outbound transfers held for human approval by default. Read: Most coding benchmarks score edits to a repository. Isu-bench scores profiling, rewriting, deploying and surviving a reboot across three servers, so its 16.6-point gap between the two leaders measures open-ended operations work, with cost and time per run. Read: On GPT-6.2 and newer models, an agent that edits or trims earlier messages loses every reasoning item after the edit and has to reason again. OpenAI says append-only histories keep that reasoning, so they get more cache hits and spend fewer tokens.