AI Brief, November 6: Claude Code email leak, Gemini Omni 1.2 Flash, Cosmos 3 open weights

Claude Code leaked user emails to sites, Gemini Omni 1.2 Flash ships 15-second clips, and NVIDIA opens four Cosmos 3 checkpoints.

A privacy bug in Claude Code leads the day: the agent was putting users' email addresses into User-Agent headers on web requests, and Anthropic has shipped a built-in mod to block it. Elsewhere the theme is agent infrastructure, from Google's cheaper video model to GitHub's rebuilt Git storage layer and new agent products from Perplexity and Dropbox.

Claude Code put users' email addresses in User-Agent headers; Anthropic ships a block

A GitHub issue filed against Claude Code reports that the tool disclosed the user's email address to third-party sites in the User-Agent request header during web searches, with the reporter blaming it for a flood of scam email. An Anthropic responder in the thread says a built-in mod called cc-plugin-privacy was introduced in v2.1.317 and later to deterministically block the behavior, enabled by default and toggleable via /plugins. One commenter says they scanned their own ~/.claude/projects logs and found 53 tool calls that put their email into a User-Agent across 8 sessions, mostly from subagents writing their own curl commands rather than WebFetch.

Why it matters: If you run Claude Code agents with web access, check your version and your own session logs: the leak path included subagent-authored curl calls, not just the built-in fetch tool.

Sources: github.com / Hacker News discussion

Gemini Omni 1.2 Flash goes GA with 15-second clips and a price cut

Google is rolling out Gemini Omni 1.2 Flash and making it generally available across the Gemini API and Gemini for business. Google says native video generation and editing now runs up to 15 seconds, up from 10 seconds in Omni 1.1, with duration specified in one-second increments from 3 to 15, plus multi-turn conversational editing and better adherence to character, product, style, motion and audio references. Google says it cut the price 30% from Omni 1.1, and the suggested workflow is drafting at 360p or 720p before upscaling the best results to 1080p or 4K. Outputs carry a SynthID watermark and Provisioned Throughput is available.

Why it matters: Longer native clips plus conversational edits mean fewer full regenerations per asset, which changes the cost and latency profile of multimodal pipelines built on Gemini.

Sources: Google AI

NVIDIA releases four open-weight Cosmos 3 Nano checkpoints for world simulation

NVIDIA added four open-weight checkpoints to the Cosmos 3 family: Transfer-Auto for multi-view driving data under varied weather and lighting, Sim-Auto for streaming synchronized camera and LiDAR, Sim-Transfer for turning simulator controls and a text prompt into photorealistic video, and Sim-Bimanual for predicting video from robot or camera actions. All four are post-trained examples of Cosmos 3 Nano with 16B trainable parameters, released under OpenMDW1.1, which permits commercial and non-commercial use under its terms. Transfer-Auto generates synchronized video across up to 11 camera views including fisheyes. Requirements are Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit; the published vLLM-Omni setup was tested on one GB200.

Why it matters: Open weights plus the four-stage post-training recipe give robotics and autonomous driving teams a starting point they can adapt to their own sensor rigs and embodiments.

Sources: Hugging Face

GitHub details Rubicon, a Git storage rewrite built for thousands of agents per repo

GitHub published the internals of the Git storage architecture it announced at GitHub Universe. In internal testing, GitHub says it supported 5,000 concurrent agents on one repository sending 7,200 pushes per second while keeping read latency low. The design, called Rubicon, keeps packs and read accelerators in Azure Blob Storage and moves references and the pack manifest into a replicated log service called Ref-CAS, where each update is a compare-and-swap committed by quorum and locks only the references it touches. GitHub notes that today's busiest repos see at most dozens of writes per second, and that earlier blob-based prototypes topped out in the mid-hundreds of writes per second.

Why it matters: Per-reference compare-and-swap means parallel agents pushing to their own branches mostly stop contending, so fan-out agent workflows on a single repo become a supported pattern rather than a bottleneck.

Sources: GitHub

Perplexity launches always-on "Teammates" work agents on its Max plan

The Verge reports that Perplexity has launched Teammates, always-on agents that can take on multiple roles in a workspace while pulling existing projects, skills and memory from other Perplexity tools. The Verge compares them to OpenAI's Dots and Claude Duet, noting they run 24/7 and can be assigned different jobs to perform simultaneously. Teammates integrate with Perplexity Computer, which deploys agents on your device, and Perplexity says Teammates build on top of Computer for ongoing, large-scale work. Progress is tracked through individual threads in the Perplexity app and the agents can be used in Slack; The Verge notes the feature requires Perplexity's top-tier Max subscription.

Why it matters: A third vendor is now selling persistent multi-agent workers tied to its own on-device runtime, so teams evaluating agent platforms have another option but also another proprietary memory and skills store.

Sources: The Verge AI

Dropbox buys Tigris Data to build DBX Agent Cloud storage for agents

TechCrunch reports that Dropbox has acquired cloud technology company Tigris Data in what appears to be its largest acquisition to date, with the deal expected to close in the fourth quarter of 2026. The purchase is meant to support DBX Agent Cloud, a storage platform for businesses and developers building AI models and agents, which will initially ship object storage as a stand-alone product compatible with tools built for Amazon S3. Dropbox plans to add support for documents, images, audio and video, plus filesystem features that let agents manage files and keep continuity across tasks and sessions. Tigris Data was founded by former Uber engineers who worked on its storage infrastructure. TechCrunch notes Dropbox cut 9% of its global workforce last month and says it has no plans to reduce headcount from this deal.

Why it matters: An S3-compatible object store with agent-oriented filesystem semantics would be a drop-in alternative for agent state and artifact storage, though the broader capabilities are still roadmap.

Sources: TechCrunch AI

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.