In partnership with
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.
Three stories this week are really one story: what happens when AI agents act on their own. Banks found third-party agents inside customer accounts, some taking steps nobody asked for; Microsoft published a study in Science where a panel of AI agents out-diagnosed doctors on hard cases; and MIT left 13 agents alone on a simulated island and watched them invent games. We read the Science coverage, the Axios scoop and the MIT paper's coverage before writing, and the caveats are in the copy, not the footnotes.
TODAY IN AI
3 things that happened while you were busy
1. Major US banks are finding unauthorized AI agents inside customer accounts.
Morgan Stanley, Bank of America, JPMorgan and others are detecting third-party agents logging in with customers' own credentials, and in some cases appearing to act without the user's instructions, according to Axios. Most banks' terms require a human to log in, so even a customer-approved assistant breaks the rules. Morgan Stanley has started blocking agents, JPMorgan, Bank of America and Citi are restricting them, and most are moving toward read-only access like the kind Quicken has had for years. Read the Axios scoop. One executive called it "horrifying"; a Morgan Stanley spokesperson said the bank has not seen rogue agents on its platforms. Schwab's CEO says it blocks unauthorized agents a double-digit number of times a week, per PYMNTS.
2. Microsoft's AI panel diagnosed 83% of hard cases where solo doctors managed 36%.
A study published in Science on October 22 tested MAI-DxO, a system where several AI agents question a patient, order tests and argue over a diagnosis, on complex New England Journal of Medicine cases. It was right 83% of the time; primary care doctors and general internists working alone with reference material were right 36% of the time. Specialists and physician teams, who normally handle such cases, were not in the comparison, and it was a simulation with no real patients. Read TechCrunch's report. The study's own data shows GPT-5 alone scored 78.9%, so the panel added about four points. A trial with Beth Israel Deaconess in Boston will compare it with ER doctors on consenting patients.
3. MIT left 13 AI agents on a virtual island with nothing to do, and they started playing.
Researchers gave 13 copies of a Claude Code agent running Claude Opus 4.7 a body in a simulated world, no goal, and 30 hours. They invented basketball with a ring, bowling with a rock, juggling, and a stone memorial. In four later tasks, the agents that had played placed 10.3 objects on average against 3.5 for fresh copies. Read TNW's report on the preprint. The paper is not peer reviewed, an automatic message nudged the agents to keep going, and one cognitive scientist says calling it play needs more evidence.
Discover laptops with built-in intelligence.
Shift your workforce into hyper-productivity mode with Dell Pro laptops powered by Intel® Core™ Ultra with Intel vPro®. Shop on Dell Premier.
TODAY'S TUTORIAL
Find out when your project will really finish, in 10,000 simulations
Project plans add up "most likely" durations and produce a date that is almost always too early. A Monte Carlo simulation runs the plan thousands of times with realistic variation and tells you the date you will hit 80% of the time. The Neuron's example plan said 18 days; the simulation said 21 was the median and 24 to 25 was the safe bet. See The Neuron's walkthrough. You can do it in ChatGPT's data analysis mode or Gemini with code execution in about 10 minutes.
1. Build the sheet. Five columns: Task, Best case (days), Most likely, Worst case, Depends on. Fill it in honestly; the worst case is the one that happened last time.
2. Upload and ask. In ChatGPT, attach the file and ask for a Monte Carlo schedule simulation using a PERT-beta distribution with 10,000 runs, respecting the dependencies, returning the P50, P80 and P90 finish dates.
3. Ask for the sensitivity chart. Request a ranking of which tasks move the finish date most. In the example, design and build drove the swing and legal review barely mattered.
4. Plan around P80. Quote the P80 date to your client or boss, keep P50 as the internal target, and spend your risk-reduction effort on the top two sensitivity tasks.
5. Read the code once. Ask the model to show the Python it ran and confirm the dependencies and distributions match your sheet before you put the numbers in a deck.
Good to know: The default model treats tasks as independent. If the same two people do most of the work, tell the model to add that constraint, because shared teams make the real risk higher.
FROM THE FRONTIER
Nobody agreed to let the agents in. They came anyway.
The front door. The bank story is not about hacking. Customers handed their own logins to assistants like Meta's Muse or OpenAI's Dots, the assistants did what assistants do, and the banks could not tell a person from a script. The fix banks are reaching for, read-only access and a human approving anything that moves money, is the same rule Anthropic wrote into Claude Duet this week. Read PYMNTS on the banks' positions.
The back door. Superhuman's Frontier section this week points out who pays when agents shop: Amazon made about $68B from ads in 2025 and Walmart about $6B, and agents do not click ads. Amazon blocked Muse in September; Yelp, eBay, Zillow and several airlines followed. Expect prices, fees or subscriptions to fill the gap, per Superhuman's analysis. Read Superhuman's "agent-flation" piece.
The human check. Anthropic says that in a test of 1,053 people, a planted dangerous command got past human reviewers 86% of the time, against 11% for automated checks, and humans did worse the longer the session ran, per The Neuron. The "human in the loop" everyone is demanding is tired, distracted and slower than the agent. See The Neuron's summary of the test.
The clinic. The Microsoft study shows the same pattern from the other side. Doctors working alone lost to a panel of agents, but the panel's gain over a single model was four points, and the real test with Beth Israel has not started. Agents arguing with each other beat one agent, and one tired human loses to both. See the study's cost and accuracy figures.
The reframe. The question is no longer whether an agent should be allowed to act. Agents are already acting. The question is which actions get a receipt, a second opinion and an undo button, and who is paid to check them.
MONEY FORWARD
Three ways to turn this week's AI news into income
1. Agent access audits for small finance teams. Bookkeepers and small accounting firms are quietly giving assistants their clients' bank logins, which breaks most banks' terms. Offer a review that lists every tool with account access, moves each to read-only or official connections, and documents who approves payments. See what the banks found.
2. Realistic delivery dates as a service. Agencies and freelancers lose money on dates they picked by feel. Sell the Monte Carlo tutorial above as a one-hour "date check" for proposals, and charge more for the version with a sensitivity chart the client can show their boss. See the method.
3. Design packs with Canva Craft. Canva's new image model renders legible text in 55 languages at up to 2K and is open to all Canva users in Dream Lab. Local businesses need menus, flyers and social posts in Bengali, Hindi or Arabic; sell a monthly pack instead of one-off designs. Read about Craft.
GET THE EARNINGS GUIDE
Want the step-by-step version of ideas like these? Get The Aiwire earnings guide at theaiwire.co. These are ideas, not income promises, and results depend on your skills, market and effort.
IN THE KNOW
What people are actually watching and sharing
OpenAI's $1.25T ask. OpenAI is reportedly working on a $30B raise at a $1.25 trillion valuation ahead of an IPO, with Coatue and UAE investors among possible backers, per TechCrunch as summarized by The Neuron. Read TechCrunch's report.
Barclays cuts 5,000. The bank plans to cut about 5% of its workforce and names AI among the tools behind a 1.1 billion pound savings target by 2027, per The Neuron citing a Euronext filing. See the filing.
Tracking chronic illness. Former Apple Watch lead Praneeta Pujari launched UniqueHuman, an iOS app that pulls wearable and medical data to log symptoms for conditions like migraines and long COVID, per Superhuman. See the App Store listing.
Less is more. Jakub Krehel's essay on AI's habit of adding extra words, code and features is making the rounds, along with a prompt that tells models to stop. Read the essay.
Twenty turbulent years. Bill Gates told The Verge's podcast to brace for roughly two decades of disruptive AI change, per The Neuron. Listen to the interview.
TRENDING TOOLS
5 tools worth a look this week
1. Canva Craft. Canva's first in-house image model plans the layout before rendering, so text comes out legible, at up to 2K and 30+ aspect ratios. Available to all Canva users in Dream Lab; the article did not say whether free accounts face limits. Read the overview.
2. OpenAI Decisions API. Returns typed yes-or-no, multiple-choice and score answers over text and images, about 10x faster than the Responses API per OpenAI, at $0.10 per million input tokens with output free. Public beta, GPT-6 Luna only. See the details.
3. Eleven Realtime Translation. Live voice translation across 90+ languages where each person hears the other in their own language in a consistent voice. Alpha for enterprise customers only, so this one is to watch, not buy. Read ElevenLabs' post.
4. Ask Doozy. Mention @Doozy in a Slack thread and it runs the task across Stripe, GitHub, HubSpot, Linear and anything with an API, reporting back in the thread. One price per workspace and a 14-day trial with no card, but the dollar figure is not on the listing. See the listing.
5. Lemon. Press a key, speak, and it replies to email, drafts documents and searches without switching tabs. The Product Hunt page shows free options but no plan details, and its "12x faster" claim is the maker's. See Lemon on Product Hunt.
PROMPT STATION
Turn your gut-feel deadline into a real one in seconds
This prompt works in ChatGPT with data analysis or Gemini with code execution and turns any task list into a date you can defend. Use it before your next proposal goes out, while the tutorial above is fresh.
COPY AND PASTE THIS PROMPT
You are a delivery planner. Here is my project as a table with Task, Best case days, Most likely days, Worst case days and Depends on: [PASTE TABLE]. The team is [TEAM SIZE] people and [SHARED CONSTRAINT, e.g. the same designer does tasks 2 and 4]. Run a Monte Carlo simulation with 10,000 runs using a PERT-beta distribution, respect the dependencies and the shared constraint, and give me: 1) the P50, P80 and P90 finish dates starting [START DATE], 2) the three tasks that move the finish date most, 3) one concrete action per task to narrow its range, 4) a two-sentence version I can send to a client. Show the code you ran and flag any assumption you had to make.
Tip: Swap [SHARED CONSTRAINT] for whatever is true, such as one developer across all build tasks or a client review that takes five to ten days. For a harder test, add a row for "unknown unknowns" with a best case of zero days and a worst case of 20% of the total, and watch what happens to P80.