DAILY DIGEST
World Labs says Atlas is now available to all users, including its Viewfinder and Chisel web apps and a developer API. It reconstructs 3D scenes from photos or generated images, renders camera paths and exports Gaussian splats or meshes for tools such as Blender and game engines.
Try the API if your project needs image-to-3D scene assets or camera-path rendering.
Google says Gemini 4 Hydrogen is rolling out globally through AI Studio and the Gemini API, at $1.50 per million input and $7.50 per million output tokens across its 1M-token context. Antigravity’s Orchestrator pairs Carbon or Argon planning with Hydrogen execution; Google reports near-Carbon Terminal-Bench 4.1 success, 24% less quota and 19% faster completion.
Test Hydrogen on routine coding tasks; compare Orchestrator’s quota savings against your own workload.
Notion says attackers who breached inference provider Fireworks accessed some transcripts stored for Notion AI’s Insights feature, affecting fewer than 1% of workspaces. Notion disabled Insights and stopped using Fireworks for open-weight models; it advises affected users to rotate credentials entered in chats during the incident window.
Check whether your workspace was affected and rotate any credentials shared in chats during the incident window.
OpenAI’s Daybreak Red Team is in early access for scoped testing of authorized web apps, vulnerability validation and patch retesting. Defense Factory scans codebases and drafts tested fixes as merge-ready pull requests using GPT-6 Cyber; engineers must review them, and fixes are not automatically deployed.
Evaluate these tools only within authorized scopes, and keep human review between generated fixes and deployment.
OpenAI says supported API workloads can now be processed and run in India, extending its earlier data-residency offering. Initial support is for GPT-6 Luna through Responses and Chat Completions APIs using an India regional endpoint, aimed at organizations with localization requirements.
If India localization is a requirement, check whether your workload qualifies for the supported endpoint.
Anthropic’s private beta lets eligible organizations run the Managed Agents service—including its harness, sandboxes, credentials and session history—in their own cloud or on premises. Requests can route to Anthropic, Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry or a custom gateway; access is requested through an Anthropic account team.
Ask your Anthropic account team about eligibility if keeping agent infrastructure in your environment matters.
IBM released experimental 3B and 8B Granite 4.1 LLMON previews, trained with markup that separates instructions from external data. IBM reports 0.02–0.1% attack success for text injection and 3.1–5.6% for tool-output injection, while warning the learned behavior is not guaranteed or a complete defense against adaptive, direct or multi-turn attacks.
Treat LLMON as an experimental layer to evaluate, not a replacement for prompt-injection safeguards.
OpenAI says its alignment team used attacker models with increased test-time compute and repeated-success criteria to generate adversarial training data, then applied reinforcement learning to improve robustness against adaptive multi-turn attacks. It reports GPT-6 Astra safely answered 91.5% of a fixed cyber-jailbreak set, versus 59% for GPT-5.6 Sol, without production safeguards.
Treat the figures as OpenAI’s test results, not a guarantee of safety in your application.
Updated 2026-11-13 13:05 UTC
A compact digest when Nick publishes. No extra noise.