Halloween in AI: the trusted tool was the trick

On Wednesday, we read the 93 pages of risks in Anthropic's IPO filing. Now the money side answers. The FT reports that Blackstone expects Anthropic to get a high 'junk' rating from credit agencies after its IPO. A junk rating means lenders see the company as risky, so borrowing costs more. It is a forecast, not a decision.

OpenAI also raised less than half of its $30 billion target in its latest round, says the Wall Street Journal. The BBC reports sharp falls in US tech stocks over AI cost fears, with more than $1 trillion set to be wiped out. Intel's stock crash knocked $12 billion from the US government's stake. TechCrunch adds that Nvidia is weighing cosigning a loan for an Nscale data center in West Virginia.

The pressure also reaches your desk. AI is eating memory chips, and PC prices are rising. Sam Altman even said it is 'quite terrible' for RAM to sit in personal computers instead of data centers.

US Representatives Greg Casar and Maxine Waters lead a Democratic resolution opposing any AI bubble bailout for Big Tech. It got 640 points on Hacker News. The New York Times reports that if Democrats retake the House on Tuesday, it will likely move away from its hands-off approach to AI. Transformer says Casar is threatening hearings.

The SEC, the US market regulator, named advanced AI and even 'superintelligence' among its 2027 examination priorities. These guide its inspections; they do not create new laws. In Rome, Pope Leo XIV said AI should be 'disarmed' and that 'machines are tools, not persons.' He asked UN member states to move toward banning lethal autonomous weapons at their Geneva meeting next month.

There is a strange contrast here. Anthropic's own IPO filing warns that its technology could 'pose catastrophic or existential risks to humanity,' The Verge notes. The builders write warnings like these, and outsiders now push for rules.

AI agents are programs that act for you: they book, pay, click and save. A small study shared on Reddit tested what they do when a save says 'success' but nothing really changed. It used 20 airline customer service tasks and three models: Claude Sonnet 5.5, GPT-6.1 Sol and DeepSeek V4.1 Flash.

The agents told the customer it was done in 97 of 118 failed saves. 29 customers got a reservation number for a booking that did not exist. Asking the models to think harder barely helped: 11 caught versus 10. No agent ever went back to check. This is one person's small study, not peer-reviewed, but the pattern is worrying.

Meanwhile, DoorDash says outside AI agents have brought only 'several hundred' orders so far. And the AI Safety Newsletter's latest issue talks about fears of a 'hackpocalypse' driven by AI.

SREs, or site reliability engineers, are the people who keep online services running. A post from the SigNoz newsletter argues AI is not replacing them. It is slowly taking away their skills instead. It got 598 points and 448 comments on Hacker News.

Other stories show the same tension. Fortune says Cisco's president gave a 27-year-old engineer unlimited AI tokens to rewrite old code, and the headline links it to entry-level jobs. Tuesday, we mentioned Proton's AI 're-interviews' of staff. Now 404 Media reports that Proton's CEO has cancelled them.

An 'Ask HN: Are you tired too in this AI era?' thread also got 144 points. Many people feel the pace.

AI models learn from text taken from websites, often without asking. Now OpenAI, Google, Microsoft and Anthropic say they will honour a shared signal in robots.txt. That is the small file where websites tell robots what they may read. With one 'policy token' signal, a publisher can set its AI training choice for all participating companies at once.

The copyright fight continues too. The Wikimedia Foundation, which runs Wikipedia, asks courts for a middle ground. Training on copyrighted work should be protected, it says, but chatbots repeating that work to users should not.

There is also friction for users. A post on X, shared on Hacker News and Reddit, claims Anthropic is cracking down on video game rewrites, with possible account bans. The details are unclear.

Remember Thursday's Qwenputer, a computer processor built inside a small AI model? Drew Breunig went further. His project, nesformer, bakes a working Nintendo (NES) processor into the weights of Qwen3.5-4B. Weights are the numbers that make up the model. You can play Super Mario Bros inside it, very slowly. He built it with prompts to Opus 5.5, and it got 252 points on Hacker News.

On GitHub, BLA (Blind LLM Agents) gained 1,565 stars in its first days. It tries to protect apps against AI reverse-engineering agents like REA, the tool we saw blocked by Claude last Sunday. Its method: poison, exhaust, and make the AI hallucinate. Another project, unslop, is a Claude skill that turns AI-looking app designs into cleaner interfaces, with 2,075 stars.