AI news daily brief: 2026-11-06
Eight stories from November 5 and 6: six have a full standalone or guide treatment already live on Pondero, and two (a GitHub preview and a Box launch) are summarized here without a separate page. The mix: five product launches, one regulatory commitment, one model release, and one application-security update.
Anthropic connects Claude to US bank accounts for read-only spending analysis
Claude can now see inside a user's bank account, but cannot touch the money in it. Anthropic launched a beta called Personal Finance on November 5, letting Pro, Max, and Ultra subscribers in the US link bank and card accounts through Plaid and ask Claude plain-language questions about balances, spending, and recurring bills, per TechCrunch. Anthropic says the hub carries no ads and those conversations are excluded from model training. The rollout follows ChatGPT's own bank-linking feature from May by six months. One gap worth flagging: unlinking a bank stops Claude from reading new transactions, but it does not erase financial figures already sitting in old chats, saved artifacts, or shared links, so anyone who shares a Claude link risks exposing balances that persist after the account is disconnected. Full story: Anthropic lets Claude read bank accounts through Plaid, but not touch the money.
Microsoft commits to an EU Data Boundary timeline for Claude in Copilot, UK excluded
This is a roadmap promise, not a shipped feature. Microsoft said on November 5 that by the end of 2026 it will extend its EU Data Boundary, the contractual commitment to store and process data inside the EU and the European Free Trade Association, to cover Claude models running inside Microsoft 365 Copilot, per Microsoft's own blog post. One day earlier, Microsoft's own Learn documentation still read that Claude models "are currently excluded" from that boundary, and that customers in the EU boundary and in the UK have Claude disabled by default. The November 5 commitment covers the EU and EFTA only; the UK is not part of it. Tenants with Claude already enabled won't need to take any action once the capability ships, Microsoft says, but it gave no firmer date than "by the end of 2026." Full story: Claude in Microsoft 365 Copilot and the EU Data Boundary: what actually changed.
GitHub's Copilot usage API now scores how often developers act on code-review comments
GitHub expanded the Copilot usage metrics API on November 5 with a copilot_code_review object that answers a question admins have only been able to guess at: does anyone actually act on Copilot's review comments? New fields total_comments_evaluated and total_comments_addressed produce an addressed rate per repository, broken down by severity, review tier, and comment type, per GitHub's changelog. A separate field, total_merged_no_human_change_request_after_review, counts merged pull requests where no person asked for changes after Copilot's first pass. A companion update the same day added fields that flag telemetry and client-version gaps, so admins can tell a quiet seat from one whose usage data simply isn't reaching GitHub. Both require the "View Copilot Metrics" permission. Pondero folded the update into its existing Copilot billing guide alongside admin tooling for GitHub Copilot. Full story: GitHub Copilot switches to usage-based billing on June 1: what to do right now.
Meta opens live video-chat avatars to every US user of its Muse AI agent
Meta's personal AI agent can now take a video call. Muse Voice, which left limited testing on November 5, animates a customizable avatar in real time using two new models, Muse Realtime Avatar and Muse Realtime Voice, that react with timed expressions and gestures, according to TechCrunch. Users can rename the avatar, change how it looks, and adjust its voice. Muse cannot see the person talking to it; it only hears audio. The rollout covers mobile, web, and Mac today, with Meta's upcoming Muse Charm wearable getting it later. Meta's chief AI officer first demoed the feature at Connect in September. Full story: Meta opens Muse's live video-chat avatar to every US user.
GitHub adds SpaceX to its secret-scanning partner program and tightens six detectors
A leaked credential that used to just trigger an alert now gets acted on automatically. GitHub's secret-scanning partnership program added SpaceX on November 5, so a starlink_public_api_v2_secret exposed in a public repo is auto-forwarded to SpaceX for revocation instead of only notifying the repo owner, per GitHub's changelog. The same release tightened false-positive filtering on six existing detectors (Alibaba, AWS, Google, OpenVSX, Pydantic Services, and Tencent), suppressing alerts on unpaired credential fragments once their companion secret has already been filtered out. SpaceX joins a partner list otherwise dominated by cloud and SaaS vendors. Full story: GitHub adds SpaceX to its secret-scanning partner program for leaked Starlink keys.
OpenAI's gated GPT-6 Cyber model generated patches for 948 Linux bugs and found 81 Firefox flaws
OpenAI's GPT-6 Cyber model, confirmed as the official name on openai.com, scored 48.5% on ExploitGym, OpenAI's benchmark for exploiting a specified vulnerability, up from 32.1% for its predecessor, a 51% relative jump, per OpenAI's launch post. In a separate automated run, the model generated validated patches for 948 Linux kernel privilege-escalation vulnerabilities at an API cost of roughly $1.2 million, and surfaced 81 previously unknown high-severity Firefox bugs that Mozilla has since started fixing, per OpenAI's research writeup. Access to GPT-6 Cyber itself is limited to a select group of approved customers inside Daybreak Red, OpenAI's most restricted tier for offensive-security research. Full story: OpenAI's GPT-6 Cyber model patched 948 Linux flaws and found 81 Firefox bugs.
GitHub ships a redesigned pull request overview page in public preview
GitHub released an opt-in public preview of a redesigned pull request overview page on November 5: a tighter merge-readiness summary, filters for unresolved conversations, newest-first activity sorting, and header quick actions for marking a PR ready or closing it, per GitHub's changelog. GitHub is soliciting feedback through a linked community discussion and calls the redesign "a starting point" rather than a finished replacement for the classic view.
Box adds an agent-first content retrieval system built on Nvidia Nemotron
Box introduced "Box Content Retrieval" at its developer conference on November 5, a feature that lets AI agents pull relevant document sections instead of ingesting whole files, built on Nvidia's open-weight Nemotron embedding models. Nvidia's VP of generative AI software, Kari Briski, told TechCrunch she expects the approach to answer queries four times faster, with 20% fewer errors and 90% fewer tokens used, per TechCrunch. Box CEO Aaron Levie said the system sends back document chunks instead of making an agent read entire files. The feature rolls out to Box Enterprise Advance customers starting early next year.
Sources
- Claude can now analyze your spending with its new Personal Finance tool: TechCrunch, November 5, 2026
- Bringing Claude models in Microsoft Copilot into the EU Data Boundary: Microsoft 365 blog, November 5, 2026
- Usage metrics API adds code review details and addressed rate: GitHub Changelog, November 5, 2026
- Meta says US users can now video chat with their Muse AI avatar: TechCrunch, November 5, 2026
- Secret scanning upgrades detectors, adds SpaceX as partner: GitHub Changelog, November 5, 2026
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents: OpenAI, November 5, 2026 A thousand fewer paths to root: OpenAI, November 4, 2026
- New pull request overview page in public preview: GitHub Changelog, November 5, 2026
- Box taps Nvidia's Nemotron for new agent-first content retrieval system: TechCrunch, November 5, 2026
