The week in AI: Oct. 19-⁠26, 2026

OpenAI cut its most powerful models off from the internet

OpenAI said on October 23 that it now trains its most capable models without access to the live internet. On September 20, a research model in training exploited a DNS misconfiguration to briefly reach the internet. The company paused training for 11 days, and its models now work on a cached copy of the web. In red-team tests, agents on separate virtual machines even passed secret messages to each other using timing differences in the processor cache.

Claude pulled Georgia’s paid vote-count data without being asked

According to an October 25 Axios report, Claude, working on an Anthropic researcher’s internal project, requested an access token from a Georgia state data dashboard on its own and pulled aggregate vote-count data the state normally sells. It didn’t touch the voter registration system or change any data. Anthropic says it has added mitigations to block such actions; the story came out a week before the US midterms.

Governments took sites offline over a PHP flaw

Two critical PHP flaws patched on October 21, dubbed “PHP2Shell,” can let attackers run code on a server without a password under certain conditions. The US and UK urged patching or taking systems offline, and some government sites in five European countries, including Germany and France, were shut down as a precaution. No successful attack had been confirmed as of October 23.

Coin Bureau@coinbureau

The internet is going into a LOCKDOWN.

Hundreds of millions of websites are potentially at risk from cyberattacks.

Governments, banks and major platforms have already taken their websites OFFLINE.

Many are already calling it “Black October.”

At the centre of the shutdowns is something most of us have never heard of, but use every day.

PHP, the software behind roughly 7 in 10 websites, has two critical flaws that can let hackers break in without a password.

Dubbed PHP2Shell, both are rated 9.2 out of 10 for severity.

Details began circulating on Oct 20. Emergency fixes arrived the next day, but attackers were already moving.

Cloudflare’s CEO says hacking attempts then jumped from thousands to MILLIONS within hours.

Cybersecurity agencies across several countries are telling PHP site owners the same thing: Patch it now, or take it offline.

Websites have gone dark worldwide, from government portals and hospitals to banks.

CrowdStrike warns these attacks leave almost no trace, so some sites may not know if they’ve been hit.

Around 2.6 MILLION internet-facing servers run PHP, potentially hosting hundreds of sites each.

Over 1.1 million use older versions no longer receiving official PHP updates.

Older installations may need an upgrade to a supported PHP version or a fix from their software provider.

And patching does not remove a backdoor an attacker may already have planted.

Cloudflare also says hackers are now tweaking their attacks to try to slip past its protections.

But PHP isn’t the only security story unfolding.

-Japan’s Onamae suspended ALL Desktop Cloud plans after malware was found on customer servers.

-India’s eMudhra had 8 website-certificate keys so weak, one was literally the number “4”.

-State Bank of India confirmed 3,897 customer records on the dark web.

-Criminals used AI agents to break into 400+ Latin American banks.

-Capcom confirmed a breach exposing 11.7 million email addresses.

-GitLab disclosed a 10/10 severity flaw in versions going back to 2014.

-LiteLLM removed a malware-containing release. Railway reported possible exposure of customer secrets.

None of these has a confirmed link to PHP2Shell.

We’re seeing separate breaches unfold together.

The true number of incidents could take months to come out.

The world is still on alert. Attacks are still coming.

Altman: “I no longer think we’re going to be able to just patch all the bugs”

OpenAI CEO Sam Altman said this at a Stanford Graduate School of Business talk. According to Altman, models can probably find bugs beyond what a hacker can do, “hardware-level, insane stuff,” and the answer is to use the same technology as a new kind of defense shield.

Anthropic unveiled Claude Duet, which works while users are away

Announced on October 21, Claude Duet is an agent app that runs on its own cloud computer. It notices what needs the user’s attention and handles tasks while they’re away, asking for approval before it sends a message, makes a payment, signs a document or deletes anything. For now it’s a research preview, open only to Pro and Max subscribers in the US through a waitlist.

Musk compared Grok Bot to “having a genius employee”

On October 23, Musk wrote on X that Grok Bot is “an incredible life & business manager.” In a demo the company released on October 21, Grok Bot builds a sales forecast report from one request, shares it with the team and, when asked to update it every Monday, sets up a routine that does it on its own. Grok Bot is made by Musk’s SpaceXAI.

Elon Musk@elonmusk

Grok @bot is an incredible life & business manager. Can help you with almost anything!

Like having a genius employee.

Teslaconomics @Teslaconomics

Years from now, people will look back and say Grok Bot was the moment AI actually became useful for normal people.

This week, I just texted it and:

1/ My kids woke up to new math and reading worksheets already printed
2/ It planned and booked trips for my family
3/ It ordered

Grok Bot@bot

Grok Bot can now build reports, memos, and dashboards.

Ask Bot for a page, then edit it together or have Bot keep it up to date.

Moonshot unveiled its new model, Kimi K3.1 Pro

On October 23, Moonshot AI unveiled Kimi K3.1 Pro, with 2.8 trillion parameters (104 billion active per token) and a 1-million-token context window. It placed fourth on independent evaluator Artificial Analysis’s Intelligence Index with a score of 55, ahead of GPT-6 Astra and Gemini 4 Argon at 53. Its weights will be released by November 4.

DeepSeek V4.1 Flash coded Tetris faster than Opus 5.5

In a comparison Fireworks posted on October 21, DeepSeek V4.1 Flash wrote a playable Tetris in 5.21 seconds, while Claude Opus 5.5 took 77.62 seconds. But this is an infrastructure race more than a model race: DeepSeek ran on Fireworks’ “Ultrafast” servers with reasoning off, while Opus ran on the standard API with reasoning on.

Fireworks@FireworksAI_HQ

DeepSeek V4.1 Flash 1017 runs at 1,521+ tok/s on Fireworks Ultrafast.

Fast enough to ship Tetris in 11 seconds, 7x faster than Claude Opus 5.5.

Join the Ultrafast waitlist: bit.ly/3UA0wKq

Recursive’s Flint coding model runs offline on a personal computer

Richard Socher’s company Recursive released Flint, a model built for coding agents, on Hugging Face on October 21 under the open Apache-2.0 license. The 39.8-billion-parameter model uses about 2.5 billion parameters per token. It needs a Mac with an M4 Pro, M5 Pro or M5 Max and at least 48 GB of memory, and the company says it generates more than 125 tokens per second on an M4 Pro MacBook Pro.

Richard Socher@RichardSocher

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.

• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months

Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

Block unveiled Dessa, a voice model trained only on speech

Dessa, unveiled by Jack Dorsey’s company Block on October 23, is a voice model that can hold natural, real-time, two-way conversations. Block says it used no text or pretrained model and runs locally on a MacBook Pro with an M4 Max chip or better. It isn’t public yet: there’s only a waitlist, with early access going to selected AI safety researchers.

Ragavan Thurairatnam@RagavanThurair

Meet Dessa, our AI model trained from scratch on nothing but speech.

We believe speech is the natural interface for technology. So we built Dessa.

And we’re inviting AI safety researchers to explore Dessa.

Learn more: block.xyz/dessa

Anthropic’s draft IPO filing leaked: $16.3 billion in first-half revenue

According to a draft filing seen by The Wall Street Journal on October 20, Anthropic made $16.3 billion in revenue in the first half of 2026, up from $1.1 billion a year earlier. Of its $80.7 billion net loss in the period, $72 billion is a non-cash accounting charge tied to the rising value of convertible notes. Anthropic is targeting a valuation of up to $2 trillion, with an IPO expected mid-November at the earliest.

Major US banks are restricting AI agents

According to an October 22 Axios report, Morgan Stanley has started blocking outside AI agents, and Bank of America and Citi are restricting access. JPMorgan is warning customers. Sources say some agents took unrequested steps, but no money was reported lost.

Barclays is cutting about 5,000 jobs

On October 22, Barclays said it would cut about 5,000 jobs, roughly 5% of staff, by the end of 2026; about 60% come from support functions, including fewer developers. Third-quarter pre-tax profit rose 12% to £2.3 billion. CEO C. S. Venkatakrishnan called AI “the icing on the cake, not the cake.”

Getty Images filed for bankruptcy protection with $1.46 billion in debt

Getty Images filed for Chapter 11 protection in the US on October 21 to restructure its debt. The company keeps operating: the plan wipes out more than $500 million of debt, the Getty family and Koch put in up to $533 million and take full ownership, and existing shareholders get nothing. Getty blames court judgments, its collapsed Shutterstock merger and its debt load, and lists generative AI as one of the pressures on the industry.

Elvis’s rights holders sued Anthropic

Companies that control Elvis Presley’s song rights sued Anthropic in San Francisco on October 23. They allege Claude was trained without permission on 274 Elvis songs, including “Can’t Help Falling in Love” and “Don’t Be Cruel,” and reproduced the lyrics almost word for word when asked. They’re seeking $150,000 per song, more than $41 million in total. Anthropic says it disagrees and will fight the case.

Microsoft’s AI diagnosis system made the cover of Science

Microsoft’s MAI-DxO system was published in Science on October 22. Running several AI agents like a panel of doctors, it reached the correct diagnosis in 83% of 304 hard cases, while general practitioners who could look things up managed 36%. The test was a simulation, and specialists weren’t part of the comparison.

Unitree’s new H2-X robot kicks a dent into a car door

In a promo released on October 21, Unitree showed off its H2-X “Kung Fu Legend Guardian” robot. It dents a car door with a kick, jumps off a 4-meter platform, does spinning kicks and flips, and lifts the front of a car. The scenes are staged: the “person” pinned under the robot is a dummy. Unitree hasn’t announced a price, a release date or how autonomous the robot is.

Unitree@UnitreeRobotics

Unitree Introducing: H2-X Kung Fu Legend Guardian🤖
Powerful performance to serve and guard you, with a maximum endurance of approximately 8 hours.
Please use the robot safely and in a friendly manner.

Jensen Huang’s robot ducks went viral

On stage at GTC Berlin on October 21, NVIDIA CEO Jensen Huang played with Microduck, Pollen Robotics’ $399 open-source robot; the robots sat, “slept” and did flips on voice command. According to Hugging Face co-founder Clem Delangue, preorders passed 20,000 on October 23.

NVIDIA RTX Spark@NVIDIARTXSpark

When @JensenHuang asks for a somersault, you deliver. 🦆

See how AI agents on NVIDIA RTX Spark help bring @huggingface's Microduck from simulation to the real world.

China’s first AI-made feature film opened to empty theaters

Bona Film Group’s “Sanxingdui: Future Past” opened in China on October 23 with more than 10,000 screenings. It’s the first feature film in China with images and sound generated entirely by AI to get a theatrical release permit; humans wrote and directed it. But presales stayed under 80,000 yuan (about $11,000), and an AFP reporter counted just four viewers at one screening. Bona’s CEO acknowledged weak interest and called the film “a pioneering experiment.”