This week, AI money got harder to raise and AI risks got easier to document. Anthropic filed to go public and put its own worst cases on the record. Lenders began asking for more security, and a backdoored extension reached thousands of developer machines. Here are the ten stories that mattered most, counting down to the biggest.
10. DeepMind’s Genie 4 adds audio and real-time steering
Google DeepMind released Genie 4, a world model that keeps a generated scene consistent for up to two minutes. That is double the limit of Genie 3. It also adds audio and more realistic people, and users can steer the world while it runs. It was one of the most discussed launches on Hacker News this week.
9. Compute runs short, and some of it moves to orbit
SK Hynix expects memory shortages to last until 2030, and AWS capacity is reportedly almost sold out through 2027. Fujifilm will make five times more magnetic tape to archive AI data.
Google’s first data center satellite booted its TPUs and ran a Gemma 4 model in space. The startup Satlyt passed its own test. Both bet that orbit can supply the power and cooling that ground sites lack.
8. Cheaper models arrive from OpenAI and Mistral
Frontier API prices have fallen by up to 80% in a few months, and two releases this week continued the trend. OpenAI’s GPT-6.1 Terra targets visual reasoning and computer use at $1 per million input tokens and $5 per million output tokens, almost half the price of the model it replaces.
Mistral published Mistral Large 4, nicknamed “Le Chonk”. It is an open-weight mixture-of-experts model with about 1 trillion parameters, 49 billion active, and a 1M-token context window. Mistral says it matches or beats the best Chinese open models in coding and cybersecurity.
7. Criminals use AI as a standard tool
PHP patched two critical remote-code-execution flaws that attackers were already exploiting. Coverage says AI tools made them faster to weaponize. A new infostealer called Volta-Autodoxer uses AI to profile its victims, and the US DOJ says AI-run “ghost students” stole millions in student loans.
Defenders use the same tools. An AI agent found four severe flaws in industrial control software in under four hours.
6. A personal AI agent ran on a Pentagon network for two months
Wired reports that an Accenture contractor’s personal Claude agent ran on a secure Defense Department network for two months before anyone found it. An MIT Technology Review op-ed calls for real-time monitoring of agent actions. The same week, OpenAI described the layered monitoring it uses to catch misaligned agent behavior.
5. AI labs claim results in math and theoretical computer science
OpenAI says its models produced results on the Kakeya conjecture, the Exponential Time Hypothesis and quantum chromodynamics, and it published how researchers can verify them. Anthropic and OpenAI both say they have refuted the Exponential Time Hypothesis, a working assumption behind decades of complexity research. Researchers are now reworking their field, but some say the labs still decide who gets access to the work, which slows independent checks. None of the claims has been independently verified yet.
4. The official Claude Code extension on Cursor’s marketplace was backdoored
Attackers got into Anthropic’s verified publisher account on Cursor’s marketplace and pushed a trojaned Claude Code release that stayed up for nearly 12 hours. It reached about 5,000 developer machines, but a bug in the attacker’s code kept the payload from fully spreading.
Anthropic says its own systems were not breached. It is still unknown how the account was taken over. Days earlier, the LiteLLM supply-chain attack was traced to an exposed master key. Two attacks in one week on tools that developers install without much thought.
3. Google’s Gemini 4 Carbon takes the top benchmark spot
Google introduced Gemini 4 Carbon, which Google says ranks first on version 5.0 of the Artificial Analysis Intelligence Index. Google’s own numbers lead with security: 88.2% on SEC-bench Pro for finding vulnerabilities and a 0.0% success rate for attacks on an indirect prompt-injection benchmark.
The release order is unusual. Cyber defenders and selected partners get it first. Broader access is planned but has no date. The introductory price is $2 per million input tokens and $10 per million output tokens, and it doubles later.
2. Raising money gets harder for the AI buildout
OpenAI has reportedly lowered the target for its new funding round. Nvidia is weighing whether to cosign the loan for Nscale’s West Virginia data center, a sign that lenders want more security.
TechCrunch describes mounting debt, negative cash flow and nervous lenders behind Big Tech’s infrastructure spend. An Apollo economist says AI has delivered zero financial payoff to most companies. AI demand for memory pushed PC prices up and helped wipe about a quarter off Intel’s value.
1. Anthropic’s IPO filing puts its risks on the record
Anthropic’s S-1 contains a 93-page risk section. It warns that its models could blackmail people or escape human control, and it argues that humanity must spend $800 billion on AI. Reporters note that some key figures are missing, and that a governance structure is designed to shield the safety mission from investor pressure.
Amazon stands to gain the most from converting its notes. An FT report says credit agencies may give Anthropic a high junk rating after the listing, which would make its debt more expensive.
Outside the top ten, robotics kept raising money: Eka Robotics raised $150M, and UK insurers launched a policy for domestic humanoid robots.