Good morning, AI builders. Claude can now read up to 24 months of your bank transactions. Anthropic's new Personal Finance beta links your accounts through Plaid, but it can't move a cent.
If you build with AI, watch this one closely: it's a live test of how much access people will hand an assistant when real money is involved.
Welcome to the 14,159 builders reading today. Forwarded this? Join here.
The 30-second brief:
Claude Personal Finance: read-only access to 24 months of transactions, in US beta
GPT-6 Cyber: 948 Linux exploits and 81 new Firefox flaws, rated Critical
Uber's agent reality check: just four rides booked by AI agents in its busiest October week
Build with AI: audit your subscriptions with Claude in 5 minutes
ANTHROPIC
Source: Anthropic, Claude help center.
The scoop: Anthropic launched Personal Finance on Thursday, a beta that links your bank and card accounts to Claude through Plaid. Claude can read up to 24 months of transactions, but it can't move, manage or invest your money.
Key details:
What it does: answers questions like "Where did my money go this month?", flags duplicate charges, fees and rising bills, and builds charts and dashboards.
The rival: OpenAI shipped a similar Plaid feature in May. Anthropic says no ads, and hub chats aren't used for training, while ChatGPT Finances follows your existing data settings, per Engadget.
Who gets it: US Pro, Max and Ultra subscribers, in a gradual rollout on web, desktop and mobile. Team and Enterprise plans are excluded, and you can link up to 16 banks.
The catch: Claude can make mistakes, so check important figures with your bank. It isn't designed to give investment advice, and investment and loan accounts show only balances and limits, with no holdings, trades, rates or terms.
Our take: Read-only is how you earn trust before asking to act. Anthropic is already inviting banks and wealth managers to talk, so the real prize is regulated money decisions, not spending charts.
OPENAI
Source: OpenAI, via The Register and The Next Web. OpenAI-reported figures.
The scoop: OpenAI released GPT-6 Cyber on Thursday, a restricted security model it rates Critical, the top cyber risk tier in its Preparedness Framework. It found 81 previously unknown high-severity Firefox flaws and built working exploits for 948 Linux kernel bugs, 588 of them previously unknown, The Next Web reports.
Key details:
The price of a bug: the Linux run cost about $1.21M in API credits, or roughly $1,279 per working exploit, per The Neuron's summary of OpenAI's report.
Access: approved Daybreak Red customers only, through the API and Codex, widening in the coming weeks.
The catch: these are OpenAI's own numbers. The model went out of scope in 0.07% of honeypot runs, and Noma Security's Chris Hughes points to an OpenAI dashboard showing under 25% of flaws fixed.
Our take: The bottleneck in security has moved from finding bugs to fixing them. If you run Linux in production, shorten your kernel patch window now, because a $1,279 exploit won't stay exclusive to approved customers forever.
BUILD WITH AI
Audit your subscriptions with Claude Personal Finance
What you'll make: a list of every subscription you're paying for, with costs, recent price rises and likely duplicates. You'll need a US Pro, Max or Ultra plan, and the rollout is gradual.
Step by step:
Open Personal Finance in Claude's side menu. If it's missing, look under More.
Select Get started, pick goals or skip, then Continue and Link accounts with Plaid. Sign in to your bank within 10 minutes, choose accounts, and give transactions a few minutes to sync.
Paste this prompt: "Which subscriptions am I paying for? List each one with its monthly cost and last charge date. Flag duplicates and anything that rose in the past 12 months."
Refine it: "Sort these by annual cost." Then check the top entries against your bank statement, since Claude can make mistakes.
Ask Claude to turn the list into a dashboard, then cancel what you don't want yourself. Access is read-only, so it can't cancel for you.
Pro move: Start with one bank. To keep financial data out of a chat, click +, then Connectors, and toggle Plaid off. Careful: unlinking your last bank permanently resets your dashboard.
UBER
Source: Uber Q3 earnings call, via TechCrunch.
The scoop: On its Q3 call, Uber said agents booked four rides and about 500 Uber Eats orders in Halloween week, out of roughly 320M trips and orders.
Booking a ride in the app takes about 50 seconds and reordering food about 90, so agents have little time to save. DoorDash reported just a few hundred agent orders last week.
CEO Dara Khosrowshahi called it "very, very early and very limited traction," but expects agents to add demand over time.
Our take: Agents won't win by replacing a 50-second tap, so build for the slow, annoying jobs where they save real time.
FROM THE CHANNEL
Sort the real GPT-7 and Gemini 4 news from the noise
Watch for our rundown of the GPT-7 "Bel" chatter, Gemini 4 Carbon and the Qwen 4 leaks, much of it still unconfirmed.
⚡ Rapid fire
🏢 Anthropic licensed AP election data in what Adweek calls its first paid deal with a digital news publisher; terms are undisclosed.
💰 Sentra raised $150M, led by Capital One Ventures, to map and limit the data that people and AI agents can reach.
🔬 Anthropic published vault particle research in which Claude helped screen about 15B predicted proteins and solve a phage vault's structure in under four months.
🏛️ India's cybercrime agency warned financial firms about criminals using AI agents, following a similar central bank warning.
👀 Unconfirmed: Google may be testing a Gemini 4 Flash codenamed Hydrogen, with X posts claiming five checkpoints at 300-500 tokens per second.
🧰 Tool drop
Microsoft Clarity Agent Activity: a free dashboard showing which AI agents visit your site and what they do there.
Ripple-65B Preview: an Apache 2.0, 65B-parameter open-weight model tuned for agentic coding.
Grok Bot calls: approve a call and it phones a US number for you, with a live transcript and takeover.
What Do You Rate Today's Newsletter?
That's the brief. See you Monday at 6am.
The World of AI team