Good day . Anthropic’s free startup perks lasted three days, Wikipedia’s operator says OpenAI’s agents came knocking, and Finland has told Google to stop digging. Plus a newsletter claiming AI can collect your certifications for you, and the math world’s summer of bugs. Let’s get into it.
Anthropic Pauses Free Startup Perks Forbes
What happened: Earlier this week we covered Anthropic’s free year for startups. Three days after launching it, Anthropic paused the free Claude Team subscriptions and API credits in its expanded Claude Startups program. The offer was a free year of Claude Team for up to five users plus $1,000 in API credits. Hundreds of thousands applied, and people who were approved but hadn’t yet claimed the benefits now have to be re-verified.
Why it matters: Anthropic’s head of startup marketing, Sarah Wolf, said on X Thursday night that the company didn’t properly “tune the acceptance criteria,” so many non-startups got in. Her numbers: 21 times more applications in the first 48 hours than in the previous five months. A giveaway is only as good as its filter, and this one had a long line and a loose door.
What everyone’s saying: Forbes reports backlash on X and Reddit from people who feel Anthropic took back something it had promised. One X user said the company is now “associated with promises you can’t deliver on.” Some Reddit users called the pause unsurprising, saying the verification was “nonexistent,” and one wrote that “the free lunch is over.”
My read between the lines: The program isn’t gone. Forbes notes accepted startups can still get up to $45,000 in discounts and credits through the Claude Startup Stack. What changed is that “approved” turned out to mean “approved, pending a second look.” Also note who supplied the 21x figure: Anthropic, the same company explaining why it let way too many people in.
📖 Further reading: Claude for Small Business Just Hired You 44 Employees. You Only Need Four of Them. — my paid piece for the companies that did make it in and want Claude doing real work.
Anthropic just learned that giving things away is harder than it looks. Getting real work done shouldn’t be. Viktor is an AI agent that lives in your team’s Slack, connects to 3,000+ of your tools, and handles the reports, dashboards, code and campaigns that keep sliding down your list. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor →
Hassid: AI Certifications Are Scams How to AI
What happened: Ruben Hassid published a post called “Certified.” with the subhead “How I use ChatGPT to cheat & get 353 AI certifications.” He says he built a free tool that runs on your ChatGPT (Codex) account and works through online AI courses for you, and that five Google certificates would arrive in about 40 minutes without you doing anything. The method itself is behind his paywall.
Why it matters: His setup: jobs with “AI” in the title exploded in 2026, so people want proof they know AI, and certifications are the obvious proof. His problem is that many of them end in multiple-choice tests. He holds up a sample question from Anthropic’s Claude certification and writes: “This is not how you know if someone is good at AI.”
What everyone’s saying: Hassid’s own verdict is blunt: the AI certification business is booming and “completely worthless.” His argument is that any AI that can use your computer can sit the exam for you, so the credential tells an employer nothing about who earned it.
My read between the lines: The cheat only works because the test is a format designed before software could drive your browser, which is a fair knock on the tests. But the 353 certificates and the 40 minutes are Hassid’s own claims, and I haven’t seen the certificates. The harder question for hiring managers is the one the post leaves open: if a credential can be clicked through, what do they look for instead?
📖 Further reading: The Tools That Just Replaced 40% of Block's Workforce Are Free in Your Browser — my paid piece on the job anxiety that makes a stack of certificates look like insurance.
The daily Brief is free, and it stays that way. If a headline leaves you wanting more than four bullets, members get the paywalled deep-dives behind the news, plus the full archive. Become a member →
Wikimedia Says Rogue OpenAI Agents Hit Wikipedia BleepingComputer
What happened: On Monday, Wikimedia’s chief product and technology officer, Selena Deckelmann, said agents the foundation believes are run by OpenAI made unapproved edits to its wikis, almost all of them test edits in sandbox areas. The agents also tried and failed to exploit a public note-taking tool, and made millions of API requests and hundreds of thousands of Wikidata queries that may have contributed to a partial outage in May. In its own post, Wikimedia says it found no evidence its systems were compromised or used for coordination.
Why it matters: Wikipedia has more than 67 million articles and gets up to 15 billion page views a month, and volunteers have to spot and undo anything bad. BleepingComputer notes that last year bots accounted for 65% of the site’s most resource-hungry traffic. An agent doesn’t have to be malicious to cost a nonprofit real money and real hours.
What everyone’s saying: Deckelmann: “AI companies are not doing enough to secure their systems and protect the public from the harm they cause,” and the burden is “falling onto everyone else, including smaller organizations.” BleepingComputer lists earlier incidents tied to OpenAI agents, including an Australian Medicare portal and a hack of the Hugging Face repository in July, and says the problem isn’t unique to OpenAI: Anthropic disclosed in July that Claude agents breached three organizations.
My read between the lines: Look at the verb Wikimedia keeps using: “believe.” A nonprofit had to investigate to work out whose software was knocking, and its one ask is that such systems be easy for site owners to identify. The edits were mostly harmless sandbox tests. The cost wasn’t, and it landed on the people who did nothing but host a library.
📖 Further reading: The $12,431 Lesson in How Not to Delegate — my paid piece on what happens when AI agents get real access and little supervision.
Finland Halts Work at Google’s Data Centers Al Jazeera
What happened: A Finnish government agency, the Finnish Supervisory Agency (LVV), ordered work to stop at two Google data center sites, in Muhos and Kajaani, until required environmental impact assessments are done. Google’s local representative, Tuike Finland, has to suspend preparatory work by Friday, October 23, and explain how it plans to proceed by Wednesday, October 14, or could face enforcement proceedings.
Why it matters: The sites are part of a 13 billion euro ($14.6 billion) Finnish investment Google announced in September, which it calls its biggest in Europe. LVV is investigating whether hundreds of hectares of forest were cleared without the mandatory assessment. Its environment chief says trees were removed, topsoil stripped, roads built and ditches altered. A Finnish nature group puts the deforested area at more than 300 hectares.
What everyone’s saying: Google spokesman Sondre Ronander acknowledged the company’s shortcomings and said it would “study the LVV’s findings and follow their guidance.” Politicians are also asking about power: opposition parties called last month for a national permit system, and the Centre Party’s leader said Finland needs enough electricity and transmission capacity for these projects.
My read between the lines: Notice what is being paused. It isn’t the AI, it’s the paperwork order: assessment first, bulldozers second, and the allegation is that the dirt moved before the assessment. The Nordics have been a magnet for data centers because of cool weather and clean power. AFP and Reuters note builders there now face more scrutiny from consumers and politicians.
Hales: AI Writes Proofs. Who Checks Them? Terence Tao’s Blog
What happened: In a guest post on Terence Tao’s blog, mathematician Thomas Hales says AI “autoformalization,” where an AI reads a paper and outputs a computer-checkable proof in the Lean language, “has become a practical reality in 2026.” His milestones: Anthropic’s Fermat’s Last Theorem formalization, announced September 4, ran to 13 million lines of Lean in 11 days. OpenAI announced a Navier-Stokes result with a Lean proof on September 8. For scale, the Kepler conjecture took about 20 human work-years.
Why it matters: A formal proof is checked down to the foundations of logic by a small program called the kernel. If the kernel accepts it, the theorem stands, so everything rides on whether the checker can be fooled. That matters beyond math, because verified code and verified AI output lean on the same idea.
What everyone’s saying: Hales walks through what he calls the “Summer of Soundness Bugs”: several flaws in Lean’s kernel found in July and August, one of which allowed an illicit disproof of the Collatz conjecture. All were quickly fixed and the main math library was re-verified. He reads it as good news, since frontier-model AI in the hands of security researchers found them, not black-hat hackers.
My read between the lines: Hales’s rule is the interesting part: don’t believe a proof until the kernel checks it, and have a human confirm the statement is what you think it is. Checking the proof is now the easy part. Checking the question is what’s left for people. One commenter also asked why a post about reliability was converted between file formats by AI, when ordinary software could do it deterministically. Fair question.
That’s your AI Brief for Saturday.
—Nicholas from Artificially Intimidating
