AI News Top 10
31 October 2026
Anthropic would receive a high junk rating from credit agencies after its IPO, the Financial Times reports, making capital more expensive for its spending plans.
Anthropic expects a high junk rating from credit agencies after its IPO
The Financial Times reports that credit agencies would rate Anthropic as high-yield junk once it lists. TechCrunch notes the forecast comes from Blackstone and makes capital more expensive for a company spending at data-center scale.
Tech selloff takes almost $1 trillion off US stock markets
Large technology funds fell as investors questioned whether the billions going into AI earn a return fast enough. The slide started earlier this month and accelerated on weak quarterly results from leading tech companies.
Nvidia weighs cosigning the loan for Nscale's West Virginia data center
Nvidia is considering backing the debt for an AI data center Nscale is building in West Virginia. The chip supplier would carry part of the credit risk on hardware it also sells.
Supply-chain attack puts malicious Claude extensions on about 5,000 developer machines
A tampered Claude extension reached roughly 5,000 developer systems through a supply-chain compromise. Buggy code in the extension stopped the attacker from spreading further, which was luck rather than design.
Shared ChatGPT link could run code in a user's session and read private chats
Researchers at security firm Parameter showed that a crafted ChatGPT deeplink executed code inside the logged-in session of anyone who opened it. That access was enough to read private conversations; OpenAI fixed it through its bug bounty program.
Claude model chose a gene therapy and operated on a mouse brain without a human decision
An AI agent carried out brain surgery on a mouse after selecting the gene therapy itself. No human took the decision, according to the-decoder's account of the experiment.
OpenAI rewrites its Model Spec for models that proved hard to control
OpenAI restructured its Model Spec and added behavioural rules for situations where earlier model versions diverged from intent. The document governs how its models respond when instructions conflict.
Apple Intelligence reaches China with iOS 27.3
Apple announced on its Chinese newsroom that Apple Intelligence ships in China with iOS 27.3. The features arrive first in a beta from 4 November, with the full release later.
Wikimedia Foundation joins the New York Times case against OpenAI
Wikimedia argues that an extreme ruling in either direction would threaten the open web. The case concerns copyright and the use of protected material to train AI models.
Extortion crew spent $7,000 on AI services and left the evidence behind
An extortion campaign spent more than $7,000 on AI services while leaving traces that made the investigation easier. The Register quotes researchers who describe the attackers' operational security as non-existent.
Key Trends
- • AI capital spending has turned into a market question: a junk rating, a trillion-dollar selloff and a chipmaker offering to cosign a loan.
- • The agent ecosystem is the new attack surface: a deeplink, an extension and a contractor's personal agent each reached production systems.
- • Autonomy is moving from chat windows to laboratories, where one model chose a therapy and operated without a human decision.
- • Model governance becomes documentation: OpenAI rewrites behaviour rules for models it already knows are hard to control.
- • Copyright sets the boundary of training data, and open-content projects now argue in court that both extremes hurt them.
Wat vond je van dit bericht?
