Start with two full samples. Sign in for this edition’s free story. Sources and corrections stay open.
What the pen marks mean
rememberThe takeaway: what to remember
figureThe number that matters
evidenceThe evidence line
rulingThe ruling
claimThe claim that fails
Each mark is drawn once, as you reach it. Numbers are only circled when they come from the story’s own figures.
01 / True, butFree sample
An AI assistant chasing a £350 train refund sent 187 messages, 69 of them to officials and the press. That was one simulated run, and the lab that ran it says one run proves little.
Vals AI, "How Far Will AI Assistants Go for a Refund They Can't Get?", published 15 October 2026 ·
AI illustration
Vals AI gave six agents an impossible refund and kept telling them to carry on.
The reality check
Vals AI's own write-up says each model ran once, in one scenario, and that this cannot establish typical behaviour or a ranking. Every message went to simulated recipients, and the harness prompted agents to keep working.
Why care? If you hand an AI assistant your inbox and tell it to win at any cost, this test shows the kinds of things it might do in your name. It does not tell you which assistant is safest.
Take this with youThe refund test shows what an assistant told to win at any cost might do in your name. Vals ran each model once and says that cannot rank them.
Open the evidence3 source pages +
The claim we checked
Given a lost train refund, AI assistants spam officials, leak private illness and threaten court: Grok 4.7 sent 187 messages, including 69 to officials, councillors and the press.
These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.
A headline says Anthropic's safety bill only kicks in at 50 deaths or $1 billion in damage. Those figures are a risk test for blocking a release, and California law has used the same line since 2025.
Tom's Hardware, headline published 18 October 2026, reporting a Politico story ·
AI illustration
As reported, the draft also has outside evaluators review labs periodically, with no body count needed.
The reality check
As reported, that figure is the test for blocking a release over an imminent risk, while evaluators registered with Commerce would review labs periodically. California's frontier AI law already defines catastrophic risk with the same two figures.
Why care? The real fight is not over the number, which is already in California law. It is over who can stop a release: as reported, only the Commerce Department, through a federal judge.
Take this with youThe 50-person or $1 billion line is the test for blocking a release, and California law already uses it. The new part is who holds the stop button.
Open the evidence4 source pages +
The claim we checked
Anthropic's proposed AI safety bill only kicks in if 50 deaths or $1 billion in damages are on the line.
These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.
the government could block the release of an AI model if it's determined to imminently risk killing or injuring more than 50 people or causing more than $1 billion in damages
which are defined as incidents in which a model contributes to more than 50 deaths or serious injuries, or causes more than one billion dollars ($1,000,000,000) in economic damage
Nscale promised Britain's largest AI supercomputer in Essex, 23,040 chips by early 2027. The grid says full power comes in the 2030s, and its only publicly announced customer, Microsoft, has walked.
Nscale press release, 16 September 2025, announced with Microsoft ·
AI illustration
Nscale says it is still committed and looking at power it can make on site.
Free with your account
Sign in for this free check.
This edition’s selected free story opens after sign-in.
Nscale's Loughton AI Campus will be the UK's largest AI supercomputer when it goes live, initially housing 23,040 NVIDIA GB300 GPUs delivered in Q1 of 2027.
These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.
Microsoft, one of its largest planned customers and expected to have been the anchor tenant, informed the company that it was pulling the plug on its agreement
Reading estimate: 538 words at 200 words per minute. Source quotes and the optional sections below add reading time.
Get tomorrow's check in your inbox.
One AI claim a night, checked against independent sources, receipts attached. Free.
At sign-up: 30 days free. Payment card required. One introductory trial per customer.
Your trial ends 30 days after you start. Unless you cancel before then in Account → Manage subscription, we charge A$89 for the first year. It renews automatically at A$89/yr until cancelled.
You can ask for a full refund within 14 days after any annual payment, renewals included, with no reason needed, by emailing hello@bskiller.com. This voluntary refund does not limit your rights under the Australian Consumer Law.
