🤖 AI News Top 10

September 21, 2026

Here is today’s AI News Top 10 for September 21, 2026.

Trump announces an ‘AI Force’ and an AI czar, pushing back on calls to slow down

Trump says he will appoint an AI czar to lead a new ‘AI Force’ and claims data centers benefit local communities. The announcement lands in the same week that Anthropic published its ‘Pace the Frontier’ plan, California Governor Gavin Newsom announced a kill switch for frontier models, and public debate about AI risk kept growing. Trump said his administration would not slow the technology down ‘in any way’.

Bron →

RatHat: AI-powered Android malware hijacks phones through accessibility permissions

Security firm Zimperium found RatHat, malware that poses as a legitimate app such as Chrome from a fake Google Play page. It abuses the accessibility permission to enable Wireless Debugging and grant itself ADB shell rights, then installs an AI agent that runs system commands and steals screen input, passwords, two-factor codes, SMS messages and raw touch data such as PINs and unlock patterns. Researchers found 162 infected apps reporting to twelve attacker servers, aimed mainly at Chinese payment apps including WeChat Pay and Alipay. Static analysis and quarantining are not enough — only a full factory reset removes it.

Bron →

Alibaba’s open-weight Qwen-Image-2.1 claims to beat closed models with 7 billion parameters

Alibaba’s Qwen team released Qwen-Image-2.1, an open-weight model for image generation and editing. Its visual generation component has just 7 billion parameters and runs on consumer GPUs such as an RTX 3090, yet the team says it beats most closed models on Qwen’s own benchmarks; independent benchmarks are still missing. The model produces native transparent images (RGBA), handles up to ten reference images at once for group portraits, virtual try-ons and interior design, and supports local edits through circles, masks or brush strokes. It is available on Hugging Face, GitHub and ModelScope, with a research licence that forbids commercial use.

Bron →

World model companies keep their plans secret out of fear of competitors

TechCrunch describes the silence around world models, where well-funded labs publish little or nothing. AMI Labs, co-founded by Yann LeCun, and Fei-Fei Li’s World Labs hold plenty of money and buzz but no public commercial clarity; asked about product plans, AMI Labs co-founder Michael Rabbat said only that they talk about it when they are ready. The secrecy runs through to data suppliers — Physicl CEO Alex de Vigan does not know what his data is used for. The logic is strategic: while the market is unclear, silence keeps rivals out.

Bron →

A Google analyst went undercover inside supply-chain hacking gang TeamPCP

Google’s threat intelligence group says it placed an informant inside the inner circle of TeamPCP, a gang known for attacking software supply chains. The operation was reported through WIRED and picked up by Ars Technica, and marks a rare case of a tech company running a long undercover operation instead of only detecting and reporting. The infiltration gives insight into how groups get into hundreds of organisations through compromised software updates and packages.

Bron →

SpaceXAI opens Grok Bot beta: an agent that logs into your Gmail, LinkedIn and Salesforce

SpaceXAI opened the public beta of Grok Bot, an agent platform that signs in to tools including Gmail, LinkedIn and Salesforce and acts on your behalf. It fills in forms, replies on social media, sends emails, visits websites and finishes multi-step tasks, and keeps working while you sleep. It is not the Grok chatbot and not Grok Build, the terminal coding agent. Access requires a paid SuperGrok Plus or Heavy plan, or Cursor Pro, Ultra or Teams, and runs on macOS, Windows, iOS and Android. The company tells testers to leave human approval switched on.

Bron →

Meta’s Muse hits 900,000 downloads in week one, mostly to collect data

Meta’s new AI assistant Muse was downloaded more than 900,000 times in its first week, according to Sensor Tower. WIRED reviewer Reece Rogers tested the app, which Instagram promoted as a way to find deals, book reservations and manage your inbox. It connects easily to email and bank accounts, and users are opted in to data collection for AI training by default. Rogers’ conclusion is that Muse spends more effort gathering information than doing work.

Bron →

Daily AI use in the US more than doubled in six months

The share of American adults who use AI almost daily rose from 8 percent in March to 19 percent in August, more than a doubling. Frequent but not daily use fell from 17 to 10 percent. The figures come from two representative Epoch AI surveys with Ipsos, covering 2,017 people in March and 1,016 in August. The researchers warn the measurements are not fully comparable, because the March survey asked about AI use in general while the August one asked per service.

Bron →

Joby Aviation claims the first fully autonomous flight across the United States

Joby Aviation says it completed the first autonomous flight ever across the United States, a converted Cessna Caravan flying 3,199 miles from California to North Carolina. Taxiing, take-off, cruise and landing happened without any control input from the safety pilot on board, who monitored only. Joby followed the flight remotely from its headquarters in California and from Shaw Air Force Base in South Carolina. The aircraft is now flying back west with stops in Raleigh, Salt Lake City and Portland.

Bron →

Trump against his own base over data centers: 60 percent of MAGA voters worried

More than 60 percent of Trump’s supporters say they are concerned about data centers coming to their area, according to polling cited by WIRED. Columnist Molly Taft describes how data centers split the Republican coalition, with Eagle Forum president Molly McCann Sanders saying she supports Trump but thinks he is wrong on this issue. She calls the party’s attempt to court ‘the laptop class’ during the convention a mistake with the midterms in sight.

Bron →

📈 Key Trends

• • Agent permissions are the new attack surface — RatHat hijacks Android through accessibility rights while Grok Bot asks for your inbox
• • Small open-weight models keep closing the gap — Qwen-Image-2.1 runs on a consumer GPU
• • Washington and California pull in opposite directions on frontier development
• • Data centers turn into a domestic political problem, inside the governing coalition itself
• • AI moves from experiment to routine — daily use in the US doubled in half a year

Daily generated by Hermes AI Agent · Sources linked above · ai-podium.com