Know
What changedGemini 4 Carbon opens to developers
Google says Gemini 4 Carbon is now available to developers, enterprises and Google AI Ultra subscribers, following an earlier cybersecurity-defender rollout. It describes Carbon as its strongest model for cybersecurity and long-horizon agent work, including software engineering; it uses more tokens for exploration, reasoning and verification. Google still recommends Gemini 4 Argon as the default for most work.
Test Carbon on complex coding or agent tasks; compare its extra token use with Argon.
Claude Meetings joins video calls in beta
Anthropic announced Claude Meetings, a visible video-call participant that follows discussion and can draft plans, map timelines on a whiteboard, answer meeting-chat questions and retrieve data from connected apps. Notes, action items and materials are available afterward in claude.ai. The beta is for Team and Enterprise plans and initially supports Google Meet.
For team workflows, check whether your plan supports the beta and assess connected-app access.
Railway reports stolen AI credentials and charges
Railway says attackers obtained customer secrets, including AI-provider credentials, in an October 21 attack and generated an estimated $740,000 in unauthorized AI usage. It estimates up to roughly 100,000 workspaces could have been exposed; that is a maximum-exposure figure, not a count of confirmed victims. Railway says it is seeking refunds and will cover charges providers do not refund.
Audit Railway secrets and provider usage, rotate exposed credentials, and set spend alerts.
DeepMind reports formal verification of Signal code
Google DeepMind says AlphaProof Code formally verified specified security properties in libsignal’s compiled AArch64 machine code, including properties of PQXDH and the Triple Ratchet. Human experts reviewed Lean specifications, while agents generated machine-checked proofs. The reported work covers those specifications and threat models—not every possible flaw or the complete Signal app.
Treat this as a promising verification workflow, not a blanket security guarantee for Signal.
MiniMax releases H3-Image weights
MiniMax released H3-Image weights on Hugging Face. The 33.1-billion-parameter model generates and edits images, supports native output up to 4096 by 4096 pixels, and uses six-step sampling. The captured report does not specify deployment requirements or licensing terms, so check those before choosing it for a hosted or commercial workflow.
Evaluate the weights against your image-quality, hardware and licensing requirements before integrating.
New York opens frontier AI registration
New York launched a state AI responsibility and research center at SUNY Binghamton to test frontier systems and develop safety and security evaluation tools. Separately, the state opened registration under the RAISE Act: large frontier AI developers operating in New York must register by January 1, 2027, and meet safety, transparency and reporting obligations.
If you develop frontier AI in New York, review whether the registration requirement applies to your organization.