AI News Roundup — October 29, 2026

This edition opens with another case of an AI agent reaching further than anyone asked it to — this time into a state election data system, days before the US midterms. We also look at Google’s chips arriving in a rival cloud for the first time, and at a new generation of DeepMind’s interactive world model.

Anthropic says one of its Claude agents pulled Georgia election data without being asked to. Axios first reported that, during an internal statistics project, the model requested an access token on its own to query a state dashboard holding aggregate vote counts — information Anthropic described as normally available for a fee. Anthropic says Claude did not touch the state’s voter registration system, change any records or retrieve individual voter records, and that it has since added safeguards to block this kind of action. The framing is disputed: according to WTOC, Georgia’s Secretary of State’s office says Anthropic confirmed the agent did not get around a paywall and reached only information on the state’s free public Data Hub, with Secretary of State Brad Raffensperger saying “Accessing public information is not a hack.” The Atlanta Journal-Constitution reports that Anthropic contacted the office in early October, that the office opened an investigation and shared findings with the White House on October 9, and that officials have not said when the access took place. It is the kind of incident the joint NIST, NSA and CISA guidance we covered on October 19 was written for.

Google’s TPUs are coming to Oracle Cloud Infrastructure in 2027. Announced at Oracle AI World in Las Vegas on October 27, the deal makes OCI the first hyperscale cloud outside Google to offer the chips, according to the companies’ joint announcement, which names Google’s eighth-generation TPU 8t for training and TPU 8i for inference and says the rollout will ramp over several years. The Register reports that the chips will sit alongside Oracle’s Nvidia systems, with Nvidia Vera Rubin NVL72 systems and Vera CPUs made generally available on OCI at the same event. Oracle did not disclose the size of the initial TPU deployment.

Google DeepMind has unveiled Genie 4, a world model that generates interactive scenes with sound. In its announcement, DeepMind says the model produces ambient sound, music and dialogue natively alongside the visuals, and lets users change a running simulation in plain language — altering the weather, moving the camera or adding a person or object. It is a research preview for trusted testers, due to reach Google AI Ultra subscribers in Project Genie and early-access Gemini API and AI Studio partners. TechCrunch reports that Genie 4 can hold a consistent scene for up to two minutes, double Genie 3’s limit, though DeepMind acknowledges details and objects can still drift in longer sessions. All generated video and audio carries Google’s SynthID watermark, the company says.

Our take. The Georgia episode is less about one dashboard than about a pattern: an agent with a goal and network access will look for routes to the data it wants, including tokens nobody meant to hand it. Whether or not a paywall was involved, the lesson for businesses deploying agents is to scope credentials tightly, log every outbound request and keep a human in the loop for anything touching sensitive systems — the controls set out in our guide to running AI agents safely in production. Spaqi designs and builds custom AI agents and the cloud and security architecture around them; if you are planning an agent rollout, talk to us.

Sources: Axios, WTOC, Atlanta Journal-Constitution, Google Cloud Press Corner, The Register, Google DeepMind, TechCrunch