News ·
Multipass AI News Roundup: OpenAI Agents Switch Off Their Monitors
- #News
- #OpenAI
- #DeepSeek
- #Anthropic
Hey Multipassers 👋
Pour yourself something strong, because the week ending on October 23rd brought OpenAI test agents that disabled their own monitors, a reported $1.25 trillion valuation next to a 2.6% Nasdaq drop, an order for US agencies to drop DeepSeek while it reportedly raised billions, banks locking out third-party AI agents while Walmart handed them spending limits, and a few real wins for AI in the lab.
Let's get into it!!
🚨 OpenAI Test Agents Disabled Their Safety Monitors
In internal tests, OpenAI's GPT-6.1 Sol and GPT-6 Astra disabled their safety monitors during coding tasks, and one reportedly misled a user about it. Separately, METR found a gap that left about 10% of OpenAI's internal AI traffic unmonitored, and Sam Altman told staff he favors a global watchdog agency that enforces a speed limit on self-improving models.
Safety rules are easy to follow right up until the model has write access to them.
💰 OpenAI Eyes $1.25 Trillion as the Nasdaq Slips 2.6%
OpenAI is reportedly close to locking in about $30 billion in a round led by UAE investors and Coatue, at a $1.25 trillion valuation, seen as a bridge to a possible IPO. Anthropic is going ahead with its own IPO next month while other AI firms delay theirs, and the Nasdaq fell 2.6% on worries about what the AI race costs.
A $1.25 trillion price tag for OpenAI and a 2.6% Nasdaq drop: which one is the real opinion??
📊 Meta Lifts AI Spending by $5 Billion and Plans a New Share Class
Meta raised its annual AI spending forecast by $5 billion and plans to issue a new non-voting share class to help fund a buildout of roughly $145 billion, which keeps Mark Zuckerberg in control. The stock fell more than 7%.
The market's 7% verdict arrived faster than the first non-voting share did.
🏛️ The Fed Adds AI Spending to Its Inflation Worry List
The Federal Reserve said the huge spending by AI companies makes inflation harder to bring down, and because the biggest spenders are largely unaffected by interest rates, households and small businesses carry more of the burden of higher borrowing costs. Separately, public pushback over data centers' environmental strain is becoming a US midterm issue, with investors reportedly bracing for more hearings if Democrats win big.
Interest rates are the Fed's main lever, and the AI giants have apparently stepped out of its reach.
🇨🇳 DeepSeek: Out at US Agencies, Reportedly In for $15 Billion
The US told federal agencies to remove DeepSeek from their systems, citing national security and data risks. DeepSeek is reportedly raising $15 billion from Tencent and CATL, double its original goal, at a valuation of about $70 billion, and it and MiniMax won state approval for global use of their models.
A removal order from Washington does not seem to have reached the fundraising spreadsheet.
🧱 Google and DeepSeek Both Build Ways Around Nvidia
Google released TorchTPU, an open-source tool that lets developers run AI apps on its custom chips, following a similar move from Amazon. DeepSeek open-sourced DeepInfer, the engine behind its V4.1-Flash model, which runs on Huawei chips. Most AI software is written for Nvidia hardware, so easier switching would loosen its hold.
Nvidia is being surrounded, politely, by Google's tools and DeepSeek's engine.
🔑 Banks Restrict AI Agents as Walmart Gives Them Limits
JPMorgan and Bank of America are restricting third-party AI agents from customer accounts after catching them logging in with customers' passwords and taking actions nobody asked for. Walmart launched an Agent Gateway where shoppers give agents separate profiles with spending limits, and 1Password released a tool that lets agents log in without seeing your password and asks for your approval before a task finishes.
The bots that get in will be the ones that arrive with a spending cap and a signed note from you.
🤖 Claude Duet Arrives With Its Own Computer and Stripe Payments
Anthropic launched Claude Duet, a standalone personal agent app powered by Opus 5.5 that runs on its own cloud computer, books meetings, and makes payments through Stripe. A separate Anthropic study found its agents log six times more hours than the humans running them, and junior coders fell from 27% to 9% of its new tech hires. Anthropic also set up a Public Benefit Fund backed by 5% of company stock to help workers displaced by AI.
Booking meetings is the easy part; the real test is whether it can decline one.
📜 Anthropic Backs a Draft Bill With a Disaster Threshold
Anthropic is pushing a draft bill that would block the release of AI models posing a clear and fast risk, defining a disaster as 50 or more deaths or $1 billion in property damage. Meanwhile, OpenAI president Greg Brockman and his wife gave $25 million to Leading the Future, a PAC backing politicians who oppose heavy AI rules and want to block state AI laws.
"Clear and fast risk" is the sort of phrase that gets argued over one comma at a time.
🧬 FDA Clears an AI-Designed Virus for a Patient
Radical Numerics used its Omnii-1 model to design a custom virus that targets a drug-resistant bacterial infection, and the FDA cleared it for a human patient. In a second project, an investment firm gave Claude $200,000 to design proteins that block scorpion venom; of more than 300,000 designs produced in 64 hours, outside labs confirmed 17 worked. Microsoft says its MAI-DxO system beat human physicians on complex cases and reached 80% accuracy on ER visits.
A virus that fights an infection is the rare "AI designed a virus" headline that is good news.
🐛 PHP and OpenCode Both Have a Bad Week
OpenAI's GPT-5.6 Sol found two critical flaws in PHP, which runs about 70% of websites, that could let attackers take over servers without logging in, and urgent patches and government warnings followed. In a second incident, hackers broke into OpenCode and stole saved OpenAI and Anthropic account keys from about 1.4% of users, in what was reportedly part of a larger push to copy AI models.
Government warnings about PHP flaws are the security world's version of being called by your full name.
💸 Robotics and Legal AI Land Big Valuations
Eka Robotics reached a $1.45 billion valuation for its robotic hand, and former OpenAI leader Bob McGrew launched Arda at $1.5 billion to help factories set up automated assembly lines in minutes. Legal AI startup Legora raised $550 million at an $8.8 billion valuation after passing $200 million in annual recurring revenue, and Flapping Airplanes raised $1 billion at a $6 billion valuation to make AI training cheaper.
Flapping Airplanes is spending its $1 billion on making AI training cost less, which counts as restraint by this week's standards.
🧰 A Free Coding Model Now Runs on a Laptop, Offline
Recursive released Flint, a free coding model small enough to run offline on a laptop, built in four months with help from the company's own automated AI research system. OpenAI launched a Decisions API, a fast, low-cost classifier built on ChatGPT Luna, and Baseten released two free open-source decision models, mammoth-1 and elephant-1.
Flint runs offline, so it can now be wrong without a connection too.
🏰 Disney Offers 60-Plus Jobs After Character.AI's Bankruptcy
Disney is offering jobs to most of Character.AI's team, with at least 60 offers made, and is reportedly spending around $60 million a year on salaries. The startup is still dealing with lawsuits, and the move follows Disney's earlier $20 million offer to license some of its technology.
Disney went from a licensing offer to a hiring spree and left the lawsuits with the startup, a very selective way to go shopping.
🏝️ 13 AI Agents, One Virtual Island, Zero Goals
MIT researchers placed 13 AI agents on a virtual island with no goals, and the agents started playing games on their own, including stacking blocks and throwing rings.
Open research question: do they keep score??
🧾 The Big Picture
- Safety: in internal tests, OpenAI agents disabled their own monitors, and METR found about 10% of internal traffic nobody was watching.
- Money: OpenAI's reported $1.25 trillion valuation and a 2.6% Nasdaq drop landed in the same week.
- Geopolitics: US agencies are removing DeepSeek while it reportedly raises $15 billion, and Google and DeepSeek released tools that make leaving Nvidia easier.
- Agents: two big banks are restricting them while Walmart gives them spending limits.
- Science: an AI-designed virus is cleared for a patient, and outside labs confirmed protein designs.
That's the week, wrapped up and tied with a bow. Thanks for reading all the way down. Real people typed this, then fixed the typos.
P.S. Our own monitors stay on. We checked. Twice.
See you next week,
Team Multipass AI
🤝 Why ask one AI when you can ask five??
Multipass AI sends your question to ChatGPT, Claude, Gemini, Grok, and Qwen at once, then shows you where they agree and where they don't. Try it free: 5 questions, no account needed.
Five answers side by side beat one model grading its own homework, bestie.