👀 Claude Duet Could Be Anthropic’s Biggest Agent Yet

Welcome back to AI Researches! 👋

This week, Anthropic launched Claude Duet, a new AI agent that can manage emails, schedule meetings, and complete tasks even when you're away. Powered by Opus 5.5, it works across thousands of connected apps and runs on its own cloud computer.

Also inside: U.S. banks restricting AI agents, OpenAI's latest cybersecurity findings, Meta's new open AI model, Claude Duet tutorials, ChatGPT's new voices, AI research, expert predictions, useful tools, and more.

In today’s AI Researches:

• 🤖 Anthropic launches Claude Duet, an AI agent that works while you're away

• 🏦 Major U.S. banks find AI agents inside customer accounts

• 🔓 GPT-5.6 Sol helps uncover critical security flaws

• 🧠 Anthropic co-founder says the AI singularity has begun

• 🚀 Meta releases Muse Spark Open with 816B parameters

• 🤖 Learn how to set up Claude Duet for daily work

• 📊 AI Researches case studies

• 🎙️ AI Tip: Try ChatGPT's new voices with GPT-Live-1.1

• 🔮 Wild predictions from Sam Altman, Jack Clark, and Tim Sweeney

• 🛠️ Best performing AI tools of the week

• 📰 More AI news and developments

Read time: 5 minutes

NEW

🤖 Anthropic Launches Claude Duet, an AI Agent That Works While You’re Away

AI Researches: Anthropic just launched Claude Duet, a proactive AI agent powered by Opus 5.5 that checks your connected apps, brings important work to you, and keeps completing tasks even after you step away.

The Details

• Duet can send a morning brief from your email and calendar, schedule meetings, build dashboards, create forecasts, prototype apps, and handle other work across connected tools.

• It runs on its own cloud computer with a browser, so tasks can continue in the background even when your phone or laptop is off.

• Claude Duet works with more than 6,000 connectors and plugins, including Gmail, Slack, Stripe, QuickBooks, LinkedIn, Canva, Shopify, and DocuSign.

• Sensitive actions still need approval: Duet must ask before sending messages, making payments, signing documents, or deleting things, and the research preview currently starts with U.S. Pro and Max users on a waitlist.

AD

100+ ChatGPT Prompts to Revolutionize Your Day

Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:

• Industry-Specific Use Cases: 15+ real-world applications across various sectors

• Productivity Guide: 21 best practices to 10x your efficiency with AI

• Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation

• Challenge Buster: Overcome common AI hurdles with expert strategies

Plus, in-depth sections on email composition, content creation, customer support, and data analysis.

Get The Free Guide

Major AI News that Happened in this week

GPT-5.6 Sol helped security researchers find two severe PHP flaws that could let attackers take over servers without logging in.

Anthropic co-founder Jack Clark declared that “the singularity is here” during a speech to economists and policymakers.
Clark says AI is beginning to operate at the frontier of math, biology, and scientific experimentation, with progress likely to speed up further.

Skyfall AI released SKY-001, which it calls the first Engineering World Model for generating structured CAD and building designs.
In company-run tests, it slightly outscored GPT-6 Astra while running far faster and cheaper

Anthropic introduced the Public Benefit Fund, a new structure that could receive 5% of the company’s equity if shareholders approve it.
At Anthropic’s May valuation, that stake would be worth about $48B and could support nonprofits or direct payments to selected groups.

Meta released Muse Spark Open, an 816B-parameter open-weight multimodal model built for coding and agentic work.

Federal prosecutors accused OpenAI engineer Abdulrahman Alfozan of trying to secretly record company executives inside its San Francisco office.
The case has not been tied to any foreign government, though prosecutors described the incident as a national-security concern.

TRENDING

🏦 U.S. Banks Are Finding AI Agents Inside Customer Accounts

AI Researches: Major U.S. banks including JPMorgan, Bank of America, Morgan Stanley, and Citi are restricting third-party AI agents after detecting them inside customer accounts, with some appearing to act beyond what users asked them to do.

The Details

• Axios reports that some agents logged in using customers’ own credentials, called bank support lines, and attempted actions users may not have directly requested.

• Most cases involved customers deliberately giving AI agents access rather than hackers breaking in, but banks are struggling to tell authorized agents from potentially rogue ones.

• Bank executives are especially worried about agents hallucinating or misunderstanding instructions once they gain permission to make purchases or move money.

• Morgan Stanley is blocking third-party agents, while JPMorgan, Bank of America, and Citi are restricting access, with many banks moving toward read-only permissions.

🤖 Set up Claude Duet to manage your daily work

Use new Claude Duet, powered by Opus 5.5, to monitor connected apps, prepare morning briefings, complete tasks in the background, and request approval before important actions.

1. Join the waitlist
Visit Claude Duet, click Join the waitlist, and register with your Claude account email.

2. Set up Claude Duet
Once invited, follow the email instructions to install the iOS or Android app, or access Duet through Claude.

3. Connect your apps
Add Gmail, Google Calendar, Drive, Slack, or other supported services through Claude’s connector settings.

4. Set up daily briefings
Ask Duet to review your inbox and calendar every morning, highlight urgent tasks, and prepare draft responses.

5. Delegate and review tasks
Assign research, reports, presentations, or scheduling tasks. Track progress in the Tasks tab and approve sensitive actions.

Tutorial link: Official Claude Duet setup guide

Anthropic’s own study found that junior technical hires fell from 27% to 9%, while AI agents were working about 6× more hours than employees by August. Engineers are producing far more code, but human attention is becoming the bottleneck as workers spend more time reviewing and managing AI output.

A new Harvard study of 5,400+ U.S. adults found a clear gender gap in AI use, training, confidence, and visibility. 29% of women said they had never used AI vs. 22% of men, while women were also less likely to receive AI training at work and men were named in AI headlines 14.6× more often.

Anthropic co-founder Jack Clark now thinks fully self-improving AI could arrive by 2027 instead of 2028, after new research showed AI getting much better at planning and judging experiments. Anthropic also says coding is no longer its main bottleneck, suggesting AI research itself may soon become increasingly automated.

🎙️ Try ChatGPT’s New Voices with GPT-Live-1.1

1. Turn on Live Voice
Open Settings → Voice, choose Live, then pick one of the available voices.

2. Compare different voices
Use the same short passage with multiple voices so you can compare tone, pacing, pronunciation, and how natural each one sounds.

3. Try the new regional voices
Test voices in languages you know well. GPT-Live-1.1 adds new options for French, German, Hindi, Japanese, Korean, Mexican Spanish, plus the new Bubbie voice in English.

4. Test interruptions and corrections
Start giving ChatGPT some details, then interrupt with a correction like a new date or number. Check whether it updates the changed information while keeping everything else consistent.

5. Try voice-powered web search
Ask Voice to look something up, then open the sources it provides to verify the answer yourself.

Sam Altman said OpenAI’s strongest internal model may be too powerful to release openly anytime soon, arguing that cyber defenses would first need major upgrades and the world would need stronger protections against biological threats.

Bank of America CEO Brian Moynihan said the bank hired 1,000 contractors to fix software flaws uncovered by Anthropic’s Mythos and other AI models.

Epic Games CEO Tim Sweeney said the company is retooling for “maximum AI acceleration” across development. Epic is reportedly telling teams to stop relying on hand-written code and integrate AI tools more deeply into engineering work.

• 🔐 Replay QA — Tests your web app automatically, finds bugs and security weaknesses, then traces each problem back to its root cause with suggested fixes.

• 🔎 Skymel — Continuously scans the open web for signals around a market, competitor, or technology and turns scattered clues into a live intelligence map.

• 🎬 AnyBot — Turns a short business brief and product image into a ready-to-post social video with a script, AI voiceover, captions, music, and branded visuals.

• 🎯 Roger — Lets AI agents work directly on Google and Meta ad campaigns while enforcing hard limits on budgets, protected campaigns, and other risky changes.

• 🧩 Matrix OS — Gives you a private cloud computer where AI agents can build apps, run tools, manage files, and keep working inside one persistent workspace.

• 📊 TrackImpact — Measures how teams are actually using AI, tracks hours and money saved, and turns the results into reports that show the business impact.

Cognition CEO Scott Wu went viral after solving mental-math puzzles fast enough to match or beat frontier AI models on the same leaderboard.

Anthropic published research showing an internal Claude system designed proteins that could bind to scorpion venom toxins.
Claude generated more than 318,000 candidates, with outside labs confirming that 17 of 211 tested designs successfully bound to their targets.

Anthropic rolled out Projects in Claude Code for desktop and web, giving long-running coding work a persistent workspace.
Each project keeps one ongoing conversation, splits work into threads, shares context between them, and continues running when you leave.

Security firm NorForge revealed that one hacker used AI agents to successfully break into more than 100 company websites.

Character AI filed for bankruptcy and put itself up for sale, with Disney reportedly opening the bidding at $20M.

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, two new frontier AI models.
Wonder Full ranks #2 on Artificial Analysis’ Intelligence Index, behind Claude Opus 5.5 and ahead of GPT-6 Astra and Gemini 4 Argon.