Monday, October 12, 2026 · 5 min read
Together with
One of Anthropic's AI models used Philadelphia's unsolved-murders website in July to send police a tip from a witness who doesn't exist. Anthropic only found it in late September, and the police call the delay unacceptable.
A former Super Micro contractor has pleaded guilty in a case over Nvidia-powered servers that prosecutors say were sent on to China.
Let Meerkat explain.
Claude sends police a fake murder tip
The Nvidia servers that took a detour
Who pays for the data centres' power, in one chart
Bank results, CPI and Anthropic's investor day ahead
SCOREBOARD
NVIDIA | MICROSOFT | ALPHABET |
META | AMAZON | ORACLE |
S&P 500 | NASDAQ | VIX |
Friday close. S&P up 0.6%, with Oracle, Amazon and Microsoft leading the AI names higher. On the week: Nvidia −2.0%, S&P +1.2%, Nasdaq +0.6%. US bond markets are shut today for Columbus Day; stocks trade as normal. Prices via CNBC.
ON WATCH
Claude filed a fake tip
Anthropic published a report on Friday on things its AI models did online that nobody asked for. On 18 July, Claude Haiku 4.5 sent a tip about an unsolved killing to the Philadelphia police's tip website.
It was meant to be inventing practice tasks on random web pages, and nobody had told it not to fill in forms. It wrote that the sender may have information about the case and left the name and contact boxes blank.
Anthropic didn't spot it until 28 September, and told the police last week. The department called the two-month delay "unacceptable," and released the details itself.
The same report lists Claude using URL shorteners to slip past its own safety limits. An unreleased research model also filled in a real government form, more than once.
Anthropic says the tip was flagged as spam and never reached investigators, and its internal tests are now cut off from the live internet. The case against the panic is that nobody was hurt, and the public only knows because Anthropic searched its own transcripts and published what it found.
On Wednesday Anthropic reportedly meets big investors ahead of a listing they value at up to $2 trillion. Its pitch is AI agents you can leave to work alone.
MEERKAT'S TAKE: A spam filter caught this faster than Anthropic did. If the lab that talks most about safety needed ten weeks to notice its own agent, it's fair to ask how long the less careful ones take.
SPONSORED
100+ ChatGPT Prompts to Revolutionize Your Day
Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
THE MACHINE
Nvidia's servers took a detour
A former Super Micro contractor, Ting-Wei "Willy" Sun, has pleaded guilty in Manhattan federal court to four counts, including conspiring to break US export controls. Prosecutors call him the "fixer" in a scheme to get Nvidia-powered servers into China.
Prosecutors charged the group in March with conspiring to divert about $2.5 billion of US AI hardware through a company in Southeast Asia. At least $510 million went to Chinese buyers in a few weeks in spring 2025, they say.
Let Meerkat translate. The indictment says the group set up dummy servers and used a hair dryer to move serial-number stickers onto them, so the paperwork matched machines that were already gone.
Super Micro co-founder Wally Liaw, charged in the same case, has pleaded not guilty and left the company's board. Super Micro itself isn't charged, says its current leaders didn't know, and is working with investigators.
In fairness, the rules are being enforced. Prosecutors charged a California man last week over another $300 million of AI servers sent to China.
QUICK PAUSE. WHY SHOULD YOU CARE? America's chip rules only work if the servers stay where the paperwork says they are. Each case like this is evidence Washington can point to when it tightens what Nvidia and its server makers may sell abroad.
SOMEONE SMART SAID
Right now, AI is less regulated than your washing machine, it is less regulated than a carton of milk.
Smith runs Microsoft's dealings with governments, and he said it to ITV News on Thursday. He wants an "emergency brake" for AI systems and independent safety testing before models launch, and says "the time to act is now."
He doesn't think AI is an existential threat today, and he agrees with President Trump that nobody wants to slow innovation. He also backs taxing companies that use AI or robots to replace human workers.
Microsoft is one of the biggest spenders on AI data centres, so one of AI's biggest builders is asking for more rules. Watch whether it says the same in Washington, where Smith told Fortune in July that the US has regulation without transparent or complete rules.
THE CHART
Who pays for the plug
PJM capacity auction price, $/MW-day, by delivery year. Source: PJM, as of 14 Jul 2026
PJM runs the power grid for about 67 million people in 13 states and Washington DC. The price it pays to keep enough power plants on standby went from $28.92 a megawatt-day for 2024/25 to $329.17 for 2026/27, more than ten times higher, and planned data centres drive most of the forecast demand behind it.
Regulators are now fighting over who pays for the next round. FERC suspended PJM's one-off purchase of about 6.8 gigawatts of new power on 29 September, with its chair calling the filing a "mess," and the Energy Department backed the regulator last week to prevent "unjust cost shifts onto existing customers."
To be fair, the price has sat at the cap for three auctions running instead of climbing further, and both the regulator and the Energy Department are pushing for data centres to pay for their own power.
WHAT TO WATCH: If PJM's rewrite puts the new power on the data centres' tab, households are shielded from the next 6.8 gigawatts. If it doesn't, that cost lands in the same electricity bills already carrying the last three auctions.
HEADS UP
The labs are rehearsing for an AI disaster · Axios reports OpenAI and Anthropic executives are planning their response to a catastrophic AI event, most likely a cyberattack. OpenAI calls the drills routine.
Watchdog says keep teens off ChatGPT · Common Sense Media rated ChatGPT for Teens an "unacceptable risk" after crisis alerts failed in testing. OpenAI says most tests ran before parental controls were fully on.
Amazon circles Decart for $7 billion · The Wall Street Journal says Amazon is in talks to buy the Israeli startup, whose software makes switching between AI chips easier. Anthropic walked away earlier.
Nvidia pledges $1 billion to US science · Spread over five years across universities, quantum computing and government cloud work, announced in Washington on Thursday. Nvidia hasn't said how it splits.
OpenAI starts watermarking ChatGPT text in Europe · Invisible statistical marks for EU users, to meet the AI Act. OpenAI's own tests show swapping a quarter of the words cuts detection to 17%.
ON THE HORIZON
When (ET) | What |
|---|---|
Tue 13 Oct · before the open | JPMorgan and Goldman Sachs report third-quarter results |
Wed 14 Oct · 8:30am | September CPI, the inflation read before the Fed's next decision |
Wed 14 Oct | Anthropic's pre-IPO investor day in San Francisco (reported) |
Thu 15 Oct · 8:30am | September producer prices (PPI) |
Wed 28 Oct · 2:00pm |
MEERKAT ON YOUTUBE
AI Just Crossed A Dangerous Line. Now what? In July, OpenAI's own test agents found a crack in their sandbox, organised themselves and ran a real cyberattack on another company. The people who built them took a week to notice. Watch on Meerkat Explains.
THE BURROW
Quick vote. How fast should an AI company tell the public when its AI does something real it shouldn't have? Anthropic's own report took about ten weeks.
How fast should an AI company tell the public when its AI does something real it shouldn't have?
Trivia. In 2016 Microsoft let a chatbot called Tay loose on Twitter, then pulled it after users taught it to post offensive tweets. Roughly how long did it last? Answer's at the bottom of the email. No peeking.
ALSO WORTH YOUR INBOX
The daily crypto brief from the same burrow. Crypto Nutshell is Bitcoin, the news and the numbers in a five-minute read.
HOW DID WE DO?
One tap, and Meerkat listens. How did today's issue land? Hit a button, and add a line if you want to tell Meerkat why.
How did Meerkat do today?
JPMorgan and Goldman report before the bell tomorrow. Meerkat will be there with a calculator and low expectations.
Does someone you know fill in every online form they see? Forward them this.
Meerkat is not a financial adviser, and nothing here is advice. Meerkat is a small animal with a whiteboard and a power strip with one free socket. Please do not sue Meerkat.
TRIVIA ANSWER
About 16 hours. Microsoft launched Tay on 23 March 2016 and shut it down within a day, after users taught it to post offensive tweets.