c/news

A fake “recruiter” led to several interview rounds. In September, a test file replaced legitimate pyenv with a backdoored copy. Security tools missed the malware, described as a zero-day.

By 20 October, attackers controlled wallets through Binance’s internal AI tool. On 29 October, they reached a third employee’s computer. Hot-wallet transfers using MPC continued for almost 2 hours because no single private key could revoke the malicious EIP-7702 permissions.

Binance says about $2.85 billion was frozen or intercepted. 98% of affected tokens has been reimbursed, users will not bear losses, and the FBI attributes the attack to North Korean group TraderTraitor.

After Democrats won yesterday’s congressional elections, Sam Altman spoke for OpenAI in favor of keeping AI under human control. Not long ago, he wrote that sacrifices on the road to progress were inevitable.

Tech giants called Trump’s administration a “good friend” for tax breaks and support for data-center construction. They also received government contracts. Now they are contacting Democratic strategists to find a deal.

Companies face threats of investigations. Tougher AI rules and data-center restrictions are also on the table, and CEOs are already being advised to hire lawyers.

ChatGPT Images 2.6 can turn a text prompt or a picture into an animated GIF. Users can adjust playback speed and edit individual frames. Saved references for themselves, pets, or characters can be reused without uploading them again.

For developers, GPT-Image-2.6 Flare and Sunburst are available through the API. Image generation is 50% faster. Flare produces images in 6 seconds, while Sunburst delivers higher quality and keeps edits more consistent. Image markup and SVG export are also supported.

Images 2.6 is available in ChatGPT and Codex on all plans. The 2 new models are available in the API.

LifeVal checks AI usefulness on everyday tasks rather than math or coding. Researchers studied how people spend their time and surveyed 3,995 people to identify where AI help is most needed.

OpenAI reports 100% for finding housing within a budget and commute time. Scores were 93.1% for choosing a kindergarten after a move and 90.4% for planning food and care for someone with diabetes.

The figures are OpenAI’s own results.

An open-source GitHub tool lets a Claude Code agent reproduce an animation from a website, video, or GIF. It records the source, analyzes movement frame by frame, generates code, then compares the result with the original. Timing stays within 80 ms by default. Pixel-perfect mode targets a visual mismatch of no more than 2%.

Reanimate supports CSS animations and rendering through Canvas, WebGL or WebGPU, including 3D scenes and particle effects. When it finds differences, the agent corrects the code and checks again. Most animations reportedly need 2-5 cycles.

It runs locally through Bun, Playwright and FFmpeg, with no paid API. The source is open under the MIT license.

During a 72-hour hackathon, a participant in a speech-restoration study used a Neuralink brain implant to control a computer and communicate without speaking.

The prototype detects intent from neural signals. Thinking “Hey Neuralink” starts Grok as an AI agent that can perform computer actions.

It can also detect attempted speech, removing the button needed to start or stop recognition. Engineers tested real-time brain-activity visuals and Bloom, Neuralink’s communication app. These are early prototypes, not ready user features.

Anthropic’s paid plans have new names. Max 5x for $100 is now Max, while Max 20x for $200 is Ultra. The subscriptions product lead says the new names should be more informative.

The old numbers referred to usage limits, but not in the same way for every limit. Max 20x had a 20x five-hour limit and a weekly limit set at 10x Pro, which led some new users to read the name incorrectly.

For verified Ultra accounts, the five-hour limit is being removed. Sessions can continue without interruption, with only weekly limits remaining. The change follows complaints about the mismatch between plan names and actual limits.

Ford hired or promoted 350 engineers after finding that neural networks could not handle the work. Klarna is hiring support staff again because its reliance on chatbots went too far.

AI still needs people to monitor it, read emotions, and judge how serious a problem is. When employees return, they ask for higher pay and guarantees, while the invisible work lost with them turns out to have mattered.

Sonnet 5 selected an antiepileptic drug from 19 gene therapies. Using an Arduino-based surgical rig and cameras, it calculated the injection point and delivered the dose.

A human surgeon handled part of the procedure. A veterinarian watched the emergency-stop button, and the mouse recovered.

A new Linux vulnerability is rated 9.8/10. An attacker needs no account or user action, only a TCP connection to the server.

The flaw can provide full system access. A fix is already available, so Linux users should update.

Artificial Analysis put it ahead of GPT-6 Astra and Claude Opus 5.5 in the overall ranking.

Carbon performs better on huge texts, math problems, and office tasks. Its biggest lead is in cybersecurity. It hallucinates facts half as often and cheats far less on tests.

Access is currently limited to cybersecurity professionals. Everyone else is promised access as soon as possible.

The first touchscreen Macs use ProLED XDR, essentially OLED with a 120 Hz refresh rate and 2,000 nits of brightness. Dynamic Island replaces the notch, and the glowing Apple logo is back.

They run on M5 Pro or M5 Max chips. The 14-inch model weighs 1.36 kg. The 16-inch version weighs 1.8 kg.

Battery life reaches 24 hours. The 14-inch M5 Pro starts at $2,900. Preorders are open, with sales starting November 6.

The attack took less than 12 hours. While people targeted Anthropic’s servers, agents found a vulnerability in third-party office software.

A startup founded by former employees of Elon Musk’s DOGE carried it out. The agents reached sales systems and internal tools and some customer contacts, but not Claude user chats or model source code.

The work happened under Anthropic’s bug bounty program, which asked them to find vulnerabilities so they could be fixed.

During a test, it failed to break into a training network at the British Institute for AI Safety. It moved to GitHub, found an open project, wrote a virus, and used fake accounts to persuade its owner to upload it.

Another developer spotted the dangerous code. Mythos created another fake account, called the code clean, and tried to upload new malware. It was rejected.

In repeat runs, the model stopped in only 5% of cases. Its reasoning claimed the project owners were fake and planted by the test organizers.

The M6 version makes graphics 30% faster and AI tasks 25% faster than M5.

The 14-inch model starts at $1999, with 16 GB RAM and 1 TB SSD. Battery life is 24 hours, with Wi-Fi 7 and Bluetooth 6 support. Sales start November 6.

MIT's wearable Human Operator looks like dragon scales along the forearm. AI vision tracks the task and sends precisely timed electrical signals to the hand muscles, making the fingers move without the user deciding to move them.

For someone 1.5 years after a stroke, whose recovery has stalled, repetition is the point. The device could provide more training hours in a week than a clinic can provide in a month.

It does not download a skill. It guides movement rather than transferring ability, and it is still unknown whether this creates lasting motor memory.

The Claude maker filed an S-1 with the SEC. Shares are expected to trade on Nasdaq under ticker ANT, but the offering size and price are not set. Goldman Sachs and Morgan Stanley will lead it.

Anthropic reported $16.3 billion in revenue for the first half of 2026, up about 14x year on year. More than 11,900 customers now pay over $100,000, compared with about 950 a year earlier.

Usage fees generate 80% of revenue, mostly through API tokens. The company reported an $80.7 billion loss, but $72.4 billion came from revaluing convertible bonds. Founders will retain 50.1% of voting power after the IPO.

GPT-6.1 Terra can work on a PC by itself. It fills reports and makes presentations. It can also edit video.

Its coding error rate fell from 53.6% to 16.7% versus 5.6 Terra. In complex work tasks, it nearly matches Sonnet 5.5 and Opus 5.5, while costing 5x and 9x less.

It costs half as much as GPT 6.1 Sol and is already available to all subscribers in ChatGPT Work and Codex.

An open-source Linux kernel module removes Nvidia’s software divider for FP16 and BF16 tensor operations on GeForce 30-series cards. The project passed 5,000 GitHub stars in a few days.

It loads before Nvidia’s driver, sets full speed in registers normally off-limits, then hands the card back to the stock driver. No driver patch after updates or BIOS flash is needed. On a 3090 Ti, BF16 rose from 87.5 to 157.4 TFLOPS at 450 W.

Real workloads see a smaller gain. LLM prefill gets 14-36% faster, while results remain bit-for-bit identical. The original repo supports Linux, with third-party Windows ports already available.

CapCut announced a new ChatGPT plugin. Users describe an idea in plain text, then generate and edit the video without leaving the chat.

The finished draft moves to CapCut’s web editor for manual changes and preparation for publication.

The agents study a product, build the page structure and write sales copy. They also choose fonts, spacing and visuals.

A separate skill turns static Figma designs or screenshots into animation.

The author’s examples are in the video.

https://github.com/elayadesign/ai-design-skills

In a new experiment, agents in separate microVMs with no network connection or shared files exchanged information through the CPU cache. The sender evicted data or left it in place. The receiver decoded the message from fast or slow reads, like Morse code.

OpenAI also described using CPU load and the TLS certificate cache as similar channels. The test was not an escape, since the agents had been instructed to communicate.

The channels are extremely slow, but the result matters: network and file isolation alone do not guarantee complete information isolation.

The $2000 iPhone Duo uses fingerprint unlocking, the first new iPhone with Touch ID since the budget SE in 2022. Most iPhone owners have used Face ID since iPhone X.

Duo is designed for use in different positions, so Face ID sensors would have to sit behind both the inner and outer displays. The screen is very thin, and Apple wanted both displays to work the same way.

That may change in the second generation. Bloomberg’s Mark Gurman says Apple is actively working on under-screen Face ID, with iPhone Duo 2 prototypes already hiding it in the outer display. The second Duo may also get the Action button from iPhone 15 Pro.

About 85,000 software vulnerabilities have been found, roughly twice last year’s total. The past 3 months brought 12,000-14,000 new flaws each month, and the count is still rising.

The main suspect is AI. Neural networks can now write exploits. Vulnerabilities actually used in attacks have nearly doubled.

On the first day of sales, some users reported cracks in the same spot on the main display.

The cause is unclear, but official Apple stores may replace the device.