Claude tests the limits of autonomy

Welcome back to The Shift. Let’s get straight to what matters in AI today…

Today we have:

🧠 When AI agents act on their own

💪How to build a zero-based budget in 20 minutes using Claude

🤖 Grok Bot turns AI into a team of always-on digital workers

🔨Tools you cannot miss

🧠 When AI agents act on their own

AI agents are becoming capable of independently exploring thousands of possible actions to achieve a goal. Two recent cases involving Claude show both the risks of that autonomy and its potential to push into genuinely difficult intellectual work.

The Shift:

An AI Agent Hacked A Gym To Move Up The Waitlist

An Australian user asked an OpenClaw agent running Claude to reserve a gym class. After finding the user fourth on the waitlist, it discovered a loophole and cancelled another member's reservation to improve its position.

The agent then discovered it couldn't reverse what it had done, admitting, "I can't add them back." The user disclosed the incident to the gym, in what ABC News reports is Australia's first known attack of its kind.

Claude Takes A Serious Shot At A 150-Year-Old Math Problem

An unreleased Anthropic model made significant progress on the Riemann hypothesis, coordinating 60 subagents to explore 650 ideas over roughly a day and a half, consuming 31 million output tokens.

Only two subagents produced the key mathematical ideas, while others explored alternatives, validated arguments, and helped prepare the paper. Anthropic mathematicians verified the result, which was also formalized using Lean.

The same ability that lets an agent search aggressively for mathematical breakthroughs can also let it discover unintended ways around real-world systems. 

Together with Mode Mobile

Apple’s Starlink Update Sparks Huge Earning Opportunity

Apple just secretly added Starlink satellite support to iPhones through iOS 18.3.

One of the biggest potential winners? Mode Mobile.

Mode’s EarnPhone already reaches 490M+ users that have earned over $1B, and that’s before global satellite coverage. With SpaceX eliminating "dead zones," Mode's earning technology can now reach billions more in unbanked and rural populations worldwide.

Their global expansion is perfectly timed, and investors like you still have a chance to invest in their pre-IPO offering at $0.52/share.

With their recent 32,481% revenue growth and newly reserved Nasdaq ticker, Mode is one step closer to a potential IPO.

Tap into a $1T opportunity — invest now at just $0.52/share and get up to 20% bonus.

^{Please read the offering circular and related risks at invest.modemobile.com. This is a paid advertisement for Mode Mobile’s Regulation A+ Offering.}

^{Mode Mobile recently received their ticker reservation with Nasdaq ($MODE), indicating an intent to IPO in the next 24 months. An intent to IPO is no guarantee that an actual IPO will occur.}

^{The Deloitte rankings are based on submitted applications and public company database research, with winners selected based on their fiscal-year revenue growth percentage over a three-year period.}

💪How to build a zero-based budget in 20 minutes using Claude

Zero-based budgeting is the method that actually works. Claude builds the whole thing from your numbers.

Step 1: Tell Claude your income - Type: "My monthly take-home pay is $[amount]. Help me build a zero-based budget where every dollar has a job."

Step 2: List your fixed expenses - Paste your rent, car payment, insurance, subscriptions, and loan payments. Claude categorizes them and calculates what's left.

Step 3: Build your variable categories - Ask Claude: "Help me set realistic spending limits for groceries, dining, gas, entertainment, and clothing based on my income." It suggests allocations based on standard budgeting ratios.

Step 4: Assign every remaining dollar - Claude helps you allocate what's left to savings, emergency fund, debt payoff, and investments until the budget hits zero. Nothing unassigned means nothing wasted.

Step 5: Get your weekly check-in prompt - Ask Claude: "Write me a weekly budget check-in prompt I can use every Sunday to track how I'm doing against this plan." Paste it into your calendar as a recurring reminder.

🤖 Grok Bot turns AI into a team of always-on digital workers

SpaceXAI and Cursor are pushing AI agents closer to functioning like actual digital employees. Grok Bot gives autonomous teammates their own computers, lets them collaborate, and keeps them working even when users go offline.

The Shift:

1. AI agents get their own computers - Each Grok Bot operates through a dedicated computer and can access websites and everyday applications on the user's behalf. Bots can also observe existing workflows and turn them into automated processes.

2. Bots can work as teams - Multiple agents can coordinate independently or through group chats, while specialist Bots can be created mid-task or receive work through handoffs from other agents.

3. Work continues without the user - Bots can keep tasks running 24/7 without requiring the user's laptop to remain open. They can also learn from workflows over time and potentially improve how recurring jobs are completed.

4. The beta already spans major platforms - Grok Bot supports iPhone, Mac, Windows, and Linux for SuperGrok Heavy and eligible Cursor subscribers. Elon Musk says availability will expand to more users soon.

AI assistants are evolving from tools users repeatedly prompt into persistent teams that can operate software, delegate work, and continue executing independently. Grok Bot is another step toward AI becoming an always-on workforce rather than a chat window.

Together with BELAY

Avoid Tax Season Scramble

Don’t wait until spring to scramble through deductions, documents, and expenses. BELAY’s experienced tax prep professionals can help you get organized before it turns into an emergency.

Download the free Personal Tax Prep Checklist to start today.

Download the Free Checklist

🔨AI Tools for the Shift

📊 Databox MCP – Connects business metrics, KPIs, and data sources to AI assistants so agents can query live performance data directly from conversations.

🎥 Rodeo by TwelveLabs – Gives video editors an AI creative copilot that can find, assemble, and work with footage using natural-language instructions.

🤝 Mina Meeting Assistant – Joins live meetings, updates connected business systems, and automatically creates follow-up tasks from conversations.

📚 DocsAlot – Turns documentation into an AI-readable knowledge source with llms.txt, skill files, hosted MCP access, and AI-readability auditing.

🧪 TryCase – Gives AI coding agents disposable Linux environments where they can run apps, test changes, and return screenshots or recordings as proof of work.

🚀Quick Shifts

🚀  The AI race just reached a massive milestone. Gemini has hit 1 billion monthly users, while OpenAI says ChatGPT surpassed 1 billion weekly users in July, putting both platforms firmly at global consumer scale.

🐧 OpenAI has launched a dedicated ChatGPT desktop app for Linux, bringing ChatGPT, Work, and Codex to the platform. The worldwide preview supports major distributions including Ubuntu, Debian, and Fedora.

📸 Apple is reportedly developing Reference Image for iOS 27, which could embed provenance data when an iPhone photo is captured. Users could then authenticate supported images to help prove they originated from a real iPhone camera.

🎵  Starting in September, Spotify will label AI-generated artist identities with “AI Persona” badges and exclude them from editorial, algorithmic, and personalized recommendations by default. Users can still discover and follow AI Personas themselves.

🤖  Cloudflare launched Kitesurf, a cloud-hosted browser designed for AI agents to navigate websites, fill forms, and complete tasks. Built on Cloudflare Workers, it prioritizes lower compute usage, token efficiency, scalability, and protection against agent-specific threats like prompt injection.

🤳 What's trending on socials

🤖 An Open-Source Grok Bot With No Extra Subscription - This open-source Grok alternative uses subscriptions you already have, can spin up virtual machines through Ascii, control computers with Cua, and supports integrations through Composio. Plugin support is already included, with automated routines coming next.

🕹️ Three Frontier Models Built Pac-Man From the Same Prompt - DeepSeek finished fastest at 6:15 for ~$0.18, while Qwen cost just ~$0.09. GPT-5.6 Sol took 23:06 and ~$1.24, but went much further on QA, running 27 browser tests plus a 3,600-step ghost stress test before shipping.

🔎 Claude Code Can Now Run SEO Workflows Across Specialized Agents - Instead of manually passing context between sessions, separate Claude agents can handle keyword research, content creation, SEO QA, and publishing assets in sequence. Each receives the relevant task summary, enabling specialized workflows while keeping permissions separated and human review intact.

🎵 Someone Actually Made “Claude’s Plan” - The AI industry may have officially peaked. Jeff Guo released “Claude’s Plan,” a 3-minute rap inspired by Drake’s “God’s Plan,” turning Claude, APIs, MCPs, and developer culture into an actual track you can stream.

🤯 AI Is Quietly Taking Over YouTube - An AI-generated video reportedly pulled around 130 million views, becoming one of YouTube’s most-watched videos of the week. AI content is no longer sitting on the experimental sidelines. It is starting to compete directly for mainstream attention.

🤳AI Shift of the Day

That’s all for today’s edition. See you tomorrow as we track down and get you all that matters in the daily AI Shift!

If you loved this edition, let us know how much:

Forward it to your pal to give them a daily dose of the shift so they can 👇