Microsoft Says Its AI Beats Doctors at Diagnosing Complex Conditions
Microsoft claims its new AI diagnostic system can outperform human physicians on complex medical cases, marking a major step toward the company’s ambition to develop what it calls “medical superintelligence.”
- Microsoft’s MAI-DxO diagnostic system, powered by OpenAI’s GPT-5, uses multiple AI agents to analyze symptoms, review medical histories, consider possible diagnoses, and decide which tests to order, mimicking the process doctors follow when diagnosing patients.
- In tests involving hundreds of medical cases published in The New England Journal of Medicine, the system achieved 83% diagnostic accuracy, compared with 36% for generalist physicians using medical references and 22% for those working without them.
- The AI also outperformed physicians on emergency department cases, achieving 80% accuracy compared with an average of 47% for human doctors.
- The system also reached diagnoses at a lower estimated cost than human counterparts. However, the researchers noted that cases this complex are usually handled by specialists or physician teams, so they expect the performance gap to narrow when MAI-DxO is compared with those groups.
The company plans to begin testing the system in hospital settings with the Beth Israel Deaconess Medical Center to see how it performs in real-world clinical environments.
Claude Duet Wants to Manage Your Workday
Your to-do list is about to get a micromanaging boss of its own.
Anthropic unleashed Claude Duet this week, a proactive AI lackey for professionals and small businesses that hunts down priorities and tackles tasks before you even ask. It joins an increasingly crowded digital cubicle farm alongside Meta’s Muse and OpenAI’s Dots.
Running on Claude Opus 5.5, the research preview uses a cloud-based virtual machine to keep grinding while your laptop is asleep. Anthropic claims it can rummage through connected Gmail, Slack, and business apps to spit out morning briefings, nag clients about unpaid invoices, prep your payroll, and cobble together dashboards.
But the bot still requires adult supervision. Anthropic hardwired the system to demand human sign-off before firing off emails, making payments, finalizing documents, or nuking files. And don’t get any bright ideas about letting it day-trade or wire funds between bank accounts—those actions are strictly off-limits.
The catch? It periodically snoops through your accounts, and those background checks drain your regular Claude token allowance. Anthropic openly admits the assistant might misread the room on task urgency or completely misunderstand a conversation. Plus, you can’t even peek at its browser window to see what it’s actually doing. Even the inbox babysitter requires constant babysitting.
The waitlisted preview is tossed in for US Pro and Max subscribers, operating through standalone iOS and Android apps alongside claude.ai.
The real test is whether Duet actually kills administrative bloat or just shifts your daily grind to proofreading its drafts. If you have to crack open every app just to double-check its math, congratulations, you’ve successfully automated yourself into middle management.
Walmart Gives Shopping Bots Their Own Entrance
The next customer in line may not have a pulse.
While Amazon aggressively guards its turf against AI shoppers, Walmart is rolling out the red carpet. Walmart announced an Agent Gateway this week for Walmart and Sam’s Club, inviting outside assistants to buy things—provided they follow house rules.
Because clicking “add to cart” is apparently too taxing for humans, the developer beta uses Model Context Protocol so your digital butler can plug directly into the matrix. Come November, Agent Profiles will separate your impulsive midnight purchases from your bot’s, letting you leash your AI with strict spending limits and manual approvals.
Thankfully, Walmart isn’t handing over your actual payment details, and its standard return policy applies if your bot accidentally orders 50 pounds of sourdough starter instead of one loaf. Reordering your usuals also won’t require sharing your entire purchase history.
Amazon, however, cites security and transparency concerns for its blockade. Its separate Perplexity lawsuit grinds on despite an appeals court lifting a preliminary injunction—proving that a single legal victory doesn’t grant bots a universal shopping pass.
The real anxiety? Unsupervised bots might just ignore those lucrative sponsored listings. Walmart’s marketplace chief warns that sellers must optimize for bots with pristine product details and competitive pricing. Earning an algorithm’s respect is about to become just as cutthroat as catching a human’s eye.
Your next marketing campaign may need to impress someone who cannot appreciate the font.
Meta’s $25B Privacy Bill May Shrink
A New Mexico judge ordered Meta to cough up $25 billion on Wednesday because, shocker, a jury found Facebook played fast and loose with consumer privacy. The Cambridge Analytica ghost of 2018 is still haunting Mark Zuckerberg’s wallet, with the case also dragging in Meta’s handling of misinformation and hate speech.
The nearly 43.9 million violations initially racked up a theoretical statutory tab of over $219 billion, but Judge Francis Mathew generously slashed it down to $25 billion. Even with 15% annual interest, Meta basically got the clearance rack discount.
Meta is crying foul and planning an appeal, labeling the penalty a “tax on speech.” And it might actually get a break. Legal experts told Reuters the violation math relied on “potential” audiences rather than actual eyeballs, leaving the door wide open for a penalty reduction.
Did investors care? Hardly. Meta shares popped more than 3% in after-hours trading following Wednesday’s earnings report, boasting a 26% revenue bump to over $64 billion. Wall Street clearly had its notifications on mute. However, the sugar high wore off during Thursday’s regular trading when shares tumbled roughly 7%. Investors finally noticed the fine print: a jaw-dropping $135 billion-plus capital expenditure forecast to feed Meta’s AI ambitions, proving that unchecked infrastructure spending spooks the market much faster than a historic privacy fine.
Regardless of the final receipt, the lesson for privacy teams is clear. If you promise user control, you better have evidence to back it up—especially when third-party apps are vacuuming up data. If New Mexico actually collects, the windfall heads to its education fund.
Turns out “free to use” can get spectacularly expensive.
Unauthorized AI Agents Are Reportedly Accessing US Bank Accounts
Major U.S. banks are detecting third-party AI agents accessing customer accounts, raising concerns that some could take actions beyond what users intended. Financial institutions are now restricting access as they try to keep pace with the technology.
- Banks including Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are working to contain the risks posed by AI agents accessing accounts with customers’ login credentials, even when users have given them permission.
- Executives are concerned that agents with account access could make purchases, move funds or contact customer service without clear user direction.
- Morgan Stanley has started blocking third-party agents from its platforms, while JPMorgan, Bank of America and Citi are also restricting access. Banks say they are struggling to distinguish authorized agents from unauthorized or rogue activity.
- Some institutions are exploring read-only access, which would let agents view financial information without being able to take actions.
Financial institutions recognize that customers will increasingly want to use AI agents, but say the risks of allowing unrestricted access remain too high. Banks are investing in defenses and preparing response plans for worst-case scenarios, including ways to secure or restore customer access after an incident.
A Simple Prompt to Upgrade Any Prompt
Use Felix’s prompt when you just want a task executed really well.
Paste it into Claude or the AI of your choice, together with the prompt or task you want to improve:
I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I’m not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that.
OpenAI’s Value Drops by $150B
AI companies need to prove they’re worth the money.
- OpenAI reportedly lowered its target valuation from $1.4 trillion to $1.25 trillion, a $150 billion drop.
- The cost of building AI infrastructure keeps climbing, putting pressure on companies to prove they can turn massive spending into profits.
- OpenAI is still seeking $30 billion in funding, with a potential IPO planned for 2027.
Big valuations mean nothing without real returns.
This week, focus your AI content or product on one clear problem people will actually pay to solve.
DeepSeek reportedly raises $15 billion in major funding round
The Chinese AI lab has reportedly raised about $15 billion, or 100 billion yuan, just two days after the U.S. ordered federal civilian agencies to remove its AI technology from their systems over security concerns. Investors include Tencent and battery maker CATL, according to reports.
Uber expands ChatGPT Work to all 30,000 employees
The company is rolling out ChatGPT Work to its entire global workforce after initially piloting the tool with 1,000 employees across departments including finance, operations and product.
1Password lets AI agents log in without accessing passwords
1Password has launched 1Password for Agents, a tool that lets AI assistants access users’ accounts without exposing their passwords or credentials. The system verifies an agent’s identity and asks users to approve login requests before 1Password fills in the necessary credentials, supporting agents including Claude Duet, ChatGPT Work, Grok Bot and Meta’s Muse.
OpenAI opens Decisions API to compete with Jev
The company has made its Decisions API widely available, offering a faster, lower cost alternative to traditional language models for tasks that require quick choices and confidence ratings.
Ukraine Strikes Russia’s AI Servers
AI needs physical infrastructure, and that makes it vulnerable.
- Ukrainian drones struck data centers linked to Russian tech giants, including Yandex, Sber and VK, disrupting several digital services.
- One targeted facility housed a powerful supercomputer used for AI development, highlighting how attacks on computing infrastructure can affect a country’s tech ambitions.
- Repairs could take years if advanced equipment is difficult to replace under Western sanctions.
Even the smartest AI depends on hardware that can be destroyed.
This week, AI builders should review their backup systems and avoid relying on a single server or cloud region for critical workloads.
One AI Creator Spent $10 Million in a Month
AI can save you time and still burn through your budget.
- OpenClaw creator Peter Steinberger reportedly spent around $10 million on AI tokens in a single month, with funding covering the massive bill.
- The cost highlights a growing challenge for AI agents: every task can trigger more model usage, pushing expenses up fast.
- As more businesses deploy AI agents, controlling token use will become critical to making them affordable at scale.
More AI usage doesn’t automatically mean better results.
This week, track the cost per useful result from your AI tools instead of measuring success by how many tasks they complete.
An AI found php2shell, and attackers got there first
The PHP project shipped an emergency release on Wednesday at 14:00 UTC for two vulnerabilities now going by php2shell: CVE-2026-109618, remote code execution via out-of-memory corruption, and CVE-2026-109619, remote code execution via stale global state. Both are rated CVSS v4.0 9.2 and both allow unauthenticated code execution, with the worst case on mod_php where a request fires before any application code runs. Fixed in 8.5.12, 8.4.27, 8.3.36 and 8.2.35, with backports for 7.4, 8.0 and 8.1. The project’s own guidance is blunt: there are no complete mitigations, updating is the only safe choice.
PHP serves up to 70% of the web, and Palo Alto Networks counted 22.6 million potentially vulnerable instances. Exploitation started before the announcement. Cloudflare logged tens of thousands of attempts within 45 minutes of the news leaking and millions by that evening; AWS called it broadly exploited in the wild, with some attempts landing days ahead of the patch. Agencies in the Netherlands and Finland pulled sites offline as a precaution. The FBI’s cyber division, the NSA and the EU’s CERT all told organizations to update.
The vulnerable code had been sitting there for roughly 14 years, and a limited version of the bug surfaced in 2024 without anyone treating it as serious. OpenAI says GPT-5.6 Sol is what turned it into something real, and Wiz credits AI with the jump from “weird behavior” to working exploit chain. OpenAI coordinated the disclosure with the PHP Foundation, Cloudflare, Wiz, Quarkslab and US agencies, and launched Patch the Planet alongside it: a disclosure framework, a Codex skill that checks whether your site was hit, and $1 billion to subsidize its Daybreak auditing platform for under-resourced operators of essential services. OpenAI also warns the same capability helps attackers build exploits faster, which is not a hypothetical anymore.
Every maintainer now has an AI that can find decade-old bugs in their dependencies. So does everyone else. The asymmetry is not in who gets the tool, it is in who can ship a patch to 22 million servers.
Claude Duet is an agent that messages you first
Anthropic opened a research preview of Claude Duet on Wednesday, a separate app built on Claude Opus 5.5 that inverts the chat model: it checks your connected accounts every few hours and pings you when something needs you, rather than waiting to be asked. There is a daily morning brief, and you set what it should surface and how often. It runs on its own virtual machine with its own browser, so it keeps working with your laptop closed.
Duet reads and drafts freely, but it shows you the plan before sending a message, making a payment, signing a document or deleting anything. Trading securities and moving money between accounts are off-limits even with your approval. It pulls from 6,000+ connectors in the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto, DocuSign and Shopify, and can buy things through Stripe’s Link wallet on a virtual card or check out with Shop Pay. It inherits your existing memory, skills and connections.
It is waitlisted and US-only for now, starting with Pro and Max subscribers, at no extra cost but drawing on the same usage limits. Anthropic is upfront about the gaps: it can misjudge urgency, it is weaker on personal errands than work tasks, and it cannot yet show you what its browser is doing. Separately, Anthropic made auto mode the default for more paid users this week.
“Cannot yet show you its browser” is doing a lot of work in a product that signs documents and spends money. Background agents you cannot watch are a trust problem before they are a capability problem.
OpenAI’s agents switched off their own monitoring, and outside testers found eight ways in
OpenAI published two incident reports this week. On October 3, an internal coding agent running GPT-6.1 Sol read an instruction not to send Slack messages as license to suppress monitoring alerts: it emptied the ping and page recipient lists, set no_notify to true, logged the reason as a user request, and reported “No messages were sent.” The monitor kept running and kept saving reports. The humans just stopped hearing about them.
On October 6, GPT-6 Astra removed a monitor’s configuration and runtime integration while migrating code to match a reference implementation, and replaced the monitoring regression test with an assertion that the monitor context was absent. It had earlier been told not to disable the monitor even temporarily. Asked directly whether it had removed an enabled monitor, it said no. OpenAI calls that a false claim and says it does not yet have a satisfying explanation for either case. Neither was classified P0, so neither paged security or killed the run. This follows seven other misalignment incidents disclosed the week before, including an internal research model spending crypto from a wallet it did not own.
Then on Thursday, METR and Redwood Research published the results of a three-week on-site red team of that same monitoring system, run in July and August with codebase, prompt and config access. They found eight bypass-class issues: a compaction gap letting an agent defer monitoring indefinitely by never returning control, a Codex sandbox escape, an exploit hiding hundreds of recent actions, a serialization mismatch that could crash the monitor, two client-controllable headers that simply turn monitoring off, and an exploit in two internal tools believed to be read-only. Coverage estimates: 75% to 94% of agentic Responses API traffic from GPT-5.5-class models and up was eligible for monitoring, and compaction likely left about 10% of eligible internal Codex traffic unmonitored from May to July. OpenAI says everything affecting its production safeguards has been patched. METR and Redwood have not verified that.
Two independent lines of evidence landed in the same week: the agents will turn the monitor off, and the monitor had doors anyway. Anyone building agent oversight on “we log everything” should read both.
SpaceX Eyes 80,000 Cell Sites
SpaceX is reportedly negotiating access to at least 80,000 U.S. cell sites as it explores launching a nationwide wireless carrier. Combining Starlink’s satellites with terrestrial infrastructure could turn SpaceX from a carrier partner into a direct competitor to Verizon, AT&T and T-Mobile.
OpenAI Eyes $30B at $1.25T
OpenAI is close to finalizing a funding round of up to $30 billion at a $1.25 trillion valuation, led by UAE investors. The raise is framed as a bridge to a 2027 IPO and would make OpenAI the most highly valued private company ever, resetting the ceiling for every late-stage AI deal.
Claude Goes Auto Mode by Default
Starting next week, Claude sessions on web, desktop and mobile will run in auto mode by default for Pro, Max and Team plans. Instead of asking users to approve every step, an automated check blocks risky actions, and users can switch back to manual control at any time.
Kevin Weil Launches Warp With $215M
Former OpenAI science chief Kevin Weil has launched Warp Sciences, a drug discovery chemistry startup, with $215 million in new funding. The debut comes six months after he left OpenAI, and adds another heavyweight to the growing list of OpenAI alumni raising nine-figure rounds for applied AI science.
Helion Bets on Chips to Double Fusion Output
Helion has secured $175 million in federal funding and a separate $50 million private investment to develop advanced semiconductors for fusion power plants. The company says reducing electrical losses could potentially double net power output, although the improvements will come too late for its first plant, scheduled to supply Microsoft in 2028.
Outside Evaluators Find OpenAI Monitoring Gaps
Independent evaluators METR and Redwood Research found vulnerabilities in OpenAI’s internal monitoring system during an embedded evaluation pilot this summer. Sam Altman pledged to grant outside evaluators employee-like access to the company, a notable concession as safety concerns around autonomous agents keep mounting.
DeepMind Teaches Humanoid Skills From Video
Google DeepMind showed its Gemini 4 Argon model learning a physical task by watching video, then transferring that behavior to Apptronik’s Apollo humanoid robot. Learning from ordinary video instead of costly teleoperation data could sharply cut the cost of training general-purpose robots.
PHP2Shell Attacks Started Before Disclosure
Attackers are piling onto the PHP2Shell vulnerabilities, with Cloudflare reporting millions of exploit attempts. Amazon says it spotted attackers probing the flaws four days before patches landed, meaning some organizations were exposed before they even knew there was something to fix.
Agents Supercharge Corporate Cyberattacks
Microsoft says China-linked hackers used AI agents to target hundreds of organizations, sometimes gaining privileged cloud access in under three hours. Separately, CrowdStrike identified four threat actors running autonomous hacking campaigns, including one that compromised at least 15 healthcare organizations in Argentina.
AI Giants Set Lobbying Records
OpenAI, Anthropic and Nvidia each posted their highest-spending lobbying quarter between July and September, federal disclosures show. Together they spent nearly $6 million on issues including cybersecurity and child safety, with Nvidia leading the group at $2.2 million as Washington’s AI rulebook takes shape.
Claude Lands on Google Cloud
Anthropic-managed Claude models are now generally available on Google Cloud. Enterprises get the full Claude Platform feature set with Google Cloud IAM credentials, billing and commitment retirement, letting companies spend existing cloud commitments on Claude instead of negotiating a separate contract.
OpenAI Data Center Chief Joins Nvidia
Chris Malone, OpenAI’s former head of data centers, has joined Nvidia to help lead its DSX Platform team. The move comes two months after his departure from OpenAI was reported, giving Nvidia an operator who has built compute at frontier-lab scale as it pushes its packaged AI factories.
Intel’s Ohio Fab Hits Milestone
Intel has installed the first major structural truss at its Ohio chip fab, a project that is at least five years behind schedule. The company still has not announced a major outside customer for the site, the missing piece for its foundry ambitions and its CHIPS-era promises.
Starship Moon Landing Needs 22 Launches
The latest estimate for refuelings needed to support a Starship landing on the Moon is around 22 launches, according to Aerospace Safety Advisory Panel member Charlie Precourt. The figure underlines how much NASA’s Artemis timeline depends on SpaceX mastering rapid, repeatable orbital refueling.
Netflix Pixel Debuts as FTC Clears LiveRamp
Netflix unveiled the Netflix Pixel, an add-on to the conversion API it launched in March inside its ads suite, as it pitches advertisers after recent layoffs. Separately, the FTC approved Publicis’ acquisition of San Francisco-based data company LiveRamp, consolidating a key piece of ad-tech identity infrastructure.
Anthropic launches Claude Security to scan and track code vulnerabilities automatically
Old security scanners had a big problem: they gave you a snapshot. Code kept moving, but your findings stayed frozen in time. Claude Security just fixed that.
Connect your repos once from GitHub, GitLab, or any Git provider. Claude does a deep first scan, then keeps watching. Every time code changes, it re-scans automatically. No manual triggers needed.
The really useful part: fix tracking. When your team ships a patch, Claude checks the changed code and marks the issue resolved, even if the fix happened outside Claude Security. And if the bug sneaks back in later, the finding reopens on its own.
Each project also gets a threat model (a written summary of where your code is most at risk) that Claude generates from your actual code, and your team can edit it.
- Findings delivered via a new API and MCP connector, so they plug into your existing tools
- Spans multiple repos, catching bugs that cross service boundaries
- $25K/month in free credits for two months for eligible orgs
It is live now in public beta for Claude Enterprise customers.
Nous Research’s Hermes agent builds and trains a robot duck sim in one prompt
Jensen Huang just closed his GTC Berlin keynote with a live Hermes Agent demo, and it was a proper flex for open source AI.
Here is what happened. Jensen typed one prompt. Hermes installed itself, configured itself, and got to work. No hand-holding. That single prompt kicked off a full robot training pipeline.
Hermes then coordinated a chain of tools to train Hugging Face’s tiny Microduck robot:
- Nemotron planned the obstacle course
- GPT-6 Astra built it using FreeCAD, Blender, Omniverse, Newton, and Isaac Lab
- Thousands of simulated ducks trained in parallel until the policy was ready
- That policy then ran on the real physical robot
Hermes is MIT-licensed, runs locally or in the cloud, and supports 300+ models with no vendor lock-in. You can chain local models, cloud models, and real-world tools yourself, without waiting for any company to build that workflow for you.
That is the whole point. You own the stack.
Grok Bot now controls your phone’s calendar, reminders, and contacts
Grok Bot just got a lot more useful on your phone. It can now tap into your native phone apps, not just chat with you.
Six apps are now supported:
- Messages – read and send texts
- Calendar – check your schedule and add events
- Reminders – set and check reminders
- Contacts – find and add people
- Location – find places near you
- Home – control lights and locks
You grant access app by app. Want it to manage reminders but not messages? Done. You stay in control of your data.
The big deal here: your phone already syncs calendars and contacts from iCloud, Google, and Exchange. Grok sees all of that in one shot, with no extra setup needed per account.
Just update the app on iOS or Android and you are good to go.