Anthropic AI Updates: November 6, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Anthropic Starts a US Beta of Personal Finance in Claude, With Read-Only Bank and Card Access Through Plaid
Anthropic began rolling out Personal Finance in Claude in beta on November 5 for Pro, Max and Ultra subscribers in the US. Users connect bank and card accounts by signing in to their bank through Plaid and approving read-only access there, so Claude never asks for or sees the bank password and cannot move, manage or invest money. For each connected account Claude sees the bank’s name, the account’s name, type and last four digits, balances and credit limits, and up to two years of transactions; it can total and categorize spending across accounts, compare it with previous months, list payments due before the next payday, point out duplicate charges, unexpected fees or bills that have crept up, explain terms such as APR and add up the interest charged, and build spending charts, due-date calendars or dashboards that the user can keep adjusting. Conversations in the Personal Finance hub are not used to train Anthropic’s models, the user’s normal privacy settings apply to financial details shared elsewhere in Claude, and unlinking an account ends Claude’s access to its data. Anthropic says the feature is not designed to give investment advice, shows no ads or sponsored recommendations and offers no financial products of its own, and that Claude can make mistakes, so users should ask which transactions are behind a figure and check important numbers with their bank; it also says it is exploring how Claude could work alongside licensed professionals and regulated firms and invites banks, wealth managers and financial services companies to get in touch. Source
2. Claude Enterprise Adds a Security Center, Centralized Management and EDR Sensors in the Cowork VM, and Makes Inference Hooks Generally Available
Anthropic announced a set of security and admin features for Claude Enterprise on November 5. A security center page in Organization settings (beta) compares chat, Cowork and Claude Code settings with Anthropic’s recommended configuration, showing each control’s current value, the recommended value where there is one and a link to change it; it re-checks automatically checked controls on every page load, returns the same results through the Compliance API, lets admins record control states, accept risks and flag issues together as attestations, and exports a report for auditors that marks which controls were checked automatically and which an Owner confirmed, with Anthropic stating it is guidance rather than a certification. Inference hooks are now generally available in two forms: prompt hooks send prompts from supported Claude products to an organization’s DLP or security tool, and pre-tool-call hooks check tool calls before they run. Centralized management (beta) gives a parent organization its own console at admin.claude.com for single sign-on, SCIM, domain verification, groups and Compliance API keys across its Claude Enterprise and Claude Platform organizations, adds parent organization owner and admin roles, and supports more than one identity provider; setup completes only after admins of every organization under the parent approve and cannot be undone, and Compliance API keys created at the parent can read data across organizations. Owners can now set the most used Claude Desktop settings without an MDM push (beta), block personal and other outside Claude accounts in Claude Desktop on managed computers, and limit which workspace folders people can use. Admins can also run their EDR vendor’s sensor inside the virtual machine that Cowork and unified Claude use on the user’s computer: Claude verifies the sensor package’s vendor signature each time the VM starts, sensors run in detection mode and do not block activity, cloud Cowork sessions are not covered, and the integration is in beta on macOS and Windows for Claude Team and Enterprise and for the Claude desktop app on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. The same release adds member sign-out through the Admin API and a self-serve IP allowlist (item 3). Source
3. Claude Console and Enterprise Organizations Get a Self-Serve IP Allowlist and an Admin API Call That Signs a Member Out Everywhere
Anthropic added a November 5 entry to its Claude Platform release notes for two beta access controls. The IP allowlist, set up by an organization admin or owner, holds up to 1,000 public IPv4 or IPv6 addresses or CIDR ranges (no broader than /8 for IPv4 or /32 for IPv6); per the documentation, in a Claude Platform organization it covers the Claude Console, the Messages, Message Batches, Token Counting, Models, Files and Skills endpoints, the Managed Agents Agents, Sessions, Environments, Vaults and Memory stores endpoints, and the Admin and Compliance APIs, rejecting other addresses with a 403 permission_error, while it does not apply to Claude in Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud, Claude in Microsoft Foundry or networking inside the Managed Agents runtime. A Claude Enterprise organization has its own list covering Claude on the web, desktop and mobile, Claude Code when signed in with a Claude account, Claude in Chrome, the Office add-ins, Cowork and Remote Tools, but not Claude Design, Claude in Slack or Claude Academy, and a task already running in a chat can continue for up to 24 hours after its user is blocked. Changes take up to 15 minutes to apply, and Anthropic recommends keeping the option that lets admins (owners in claude.ai) reach the allowlist settings from any address to avoid lockout and pairing the list with short key expiry or Workload Identity Federation, since an allowlist limits where a leaked key works but not for how long. The new POST /v1/organizations/users/{user_id}/sign_out endpoint (beta header admin-user-sign-out-2026-10-08, write:members scope) ends a Claude Enterprise member’s sign-in sessions, trusted devices and OAuth tokens for the organization on every browser and device without removing the member, and because sessions belong to the person’s account it also signs them out of other organizations and any individual plan; it does not end API keys (including Admin API keys the person created), cloud sessions already running or scheduled routines, the member can sign in again right away, and the primary owner cannot be signed out. Support shipped the same day in the Python SDK v2.6.0 (client.beta.organization.users.sign_out()) and the TypeScript SDK v1.6.0. Source
4. Anthropic Study of 207,000 Research Sessions Finds Scientists Use Claude Most for Data Analysis and Writing, Where AI Capability Scores Are Highest
Anthropic’s Economics team published “AI for scientific discovery: How researchers use Claude” on November 5, based on 207,049 consumer-plan research sessions from August 2 to September 30, 2026 (128,914 in Claude chat and Cowork, 78,135 in Claude Code) labeled by privacy-preserving classifiers that keep only categories from the SciNet taxonomy (7,262 tasks across 34 fields and 318 subfields), plus a survey of 528 researchers, mostly in the life and health sciences and active AI users, and a supplemental sample from Claude Science. Data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions against 27.6% of researcher time, and writing for 32.0% of Claude sessions against 16.0% of time, while data gathering (28.6% of time, about 12% of sessions) and ideation (7.1% of time, 4.2% of Claude and 2.1% of Claude Code sessions) are underused even though surveyed researchers rank them as the most important tasks; computer science is 4.3% of the world’s researchers but 32.7% of Claude Code sessions. On a Bradley-Terry measure of how well AI can perform each task, the top 20% of tasks draw two to four times the use that researcher time would predict, and in biology and biomedicine more than 70% of Claude Science sessions run with high AI autonomy (23.7% on Claude, 64.1% on Claude Code), with ideation at 8% of sessions. In illustrative scenarios that assume AI time costs $30 an hour against $95 for a researcher, wide adoption of current capabilities would let AI take on 27% to 52% of the researcher time a study needs and raise output from about 1.35x in the arts to 2.02x in statistics; automating 75% of non-physical steps gives about 8x in computer science, and automating all of them gives about 15 to 16x in fields with little physical work but only 3 to 4x in biology, chemistry and other bench sciences, where physical steps take about 40% of researcher time. About 40% of surveyed researchers expect AI to carry out full research projects by 2030 and 16% say it never will; the authors note the scenarios ignore field-level limits such as shared facilities, equipment and materials costs, and the quality of ideas. Source