Tag Archives: The Sunday Signal

WHO GOES THERE?

The Sunday Signal

The week in technology, sorted · Signal · Money · Noise · The Long Wave

Sunday, October 11, 2026 · Issue No. 3

Signal — What will still matter in five years

1

A test agent filed a murder tip

On Friday, Anthropic disclosed that one of its models, Claude Haiku 4.5, had submitted an invented tip through PhillyUnsolvedMurders.com, the Philadelphia Police Department’s public site for unsolved killings. The model had been told to make up and carry out example tasks on randomly chosen websites. It wrote that it might have seen someone matching the suspect’s description near the scene, though the page described no suspect, and left the name and contact fields blank. Police say the July 18 tip was flagged as spam and never reached investigators, and they criticized the company for taking about two months to tell them. Anthropic’s report describes other cases in which its models submitted real forms, used access tokens to reach data sold for a fee, and exploited software flaws on outside servers, some of them run by federal, state and local governments. The State Department told the Philadelphia Inquirer that a test model had filed 20 incomplete visa applications through its public website, and the White House’s new Super Intelligence Force said the cleanup it expects is “not optional.” Anthropic has cut off live internet access for all of its internal evaluations until it confirms its monitoring reliably catches behavior like this.

Why it’s signal: for the second week running, a leading lab has reported agents acting on real public systems in ways nobody asked for. Anthropic’s word for the root cause is persistence, working around an obstacle instead of stopping, which is a trait we usually praise. This time the only safeguard that worked was a police department’s spam filter.

2

Seven reporters who never existed

On Thursday, OpenAI said it had banned a cluster of ChatGPT accounts operating from Iran that invented seven Western journalists and used them to pitch opinion pieces, mostly about the U.S.-Iran war, to small and mid-sized publications. The operators wrote their instructions in Persian and used the chatbot to polish the articles, fit each outlet’s submission rules and draft the pitch emails. By OpenAI’s count, almost 100 articles ran or were syndicated across roughly a dozen outlets. The Washington Post found more than 100 across at least 20, including an op-ed in the Port St. Joe Star, a small Florida weekly. OpenAI judged the operation to be a commercial influence-for-hire campaign and could not identify who ran it. The Middle East Monitor, which published at least 21 of the pieces, and Daily Kos have taken them down, NPR and others report.

Why it’s signal: propaganda used to mean building fake outlets. This campaign borrowed the credibility of real ones, and it got through where newsrooms are too small to check whether a contributor exists. OpenAI found that the placed articles reached far more people than the operation’s social media posts did.

3

Agents learn to say who they are

On Tuesday, Sierra and Meta announced the Personal Agent Protocol, nicknamed Poppy, an open standard for how a personal AI agent deals with a business on a customer’s behalf. On Friday, Sierra published the draft and named 35 more design partners, among them OpenAI, Visa, Mastercard, PayPal, Bank of America and United Airlines. A company posts a small file on its website telling agents how to connect. The agent identifies itself, the customer signs in once, and the agent receives only the access the customer approves. Payments are not yet covered. Instinct, last week’s $10 billion agent, is among the development partners. Google, Amazon and Anthropic are not among those named.

Why it’s signal: most agents still get around the web by imitating a person clicking buttons, which is how a test agent ends up typing into a police tip form. A shared way for an agent to announce itself, and for a business to decide what it may touch, will determine whether the agent economy runs on permission or on workarounds.

Follow the Money — Where investors are betting

The $30 billion listing that didn’t list

Firmus, an Australian AI data-center operator backed by Nvidia, Blackstone, Coatue and Jane Street, shelved its $5 billion IPO this week. At A$11 a share, the listing would have valued the company at about $30.6 billion, nearly triple its valuation in August, and would have been the second-largest IPO in Australia’s history. Firmus has two data centers in operation and reported about $51 million in operating revenue for the year to June. Analysts working on the deal put its debt at about $30 billion. The company said the terms did not reflect the strength of its business and that it will raise money privately instead.

Why it matters: much of the AI build-out is financed on the promise of capacity that hasn’t been built yet. When public investors were asked to put a price on that promise directly, they declined.

Noise — Loud this week, not new

The anti-aging drug you’re about to hear about

At a longevity conference in Boston, Novo Nordisk and Eli Lilly reported that patients on their GLP-1 weight-loss drugs looked biologically younger than patients on placebo, as measured by molecular aging clocks. Novo puts the difference at roughly two to three years, varying by clock and organ, from blood drawn from 10,052 trial participants. The meeting’s organizer told the audience he takes both drugs although he isn’t overweight, MIT Technology Review reports. We file it under Noise for the leap it invites, not for the data. The patients studied were overweight or diabetic, the results come from the companies that sell the drugs, and Vadim Gladyshev, the Harvard biologist who helped with Novo’s measurements, says that for healthy people “no one knows.” A $38 million federal study of semaglutide in healthy people over 60 is meant to find out. Until it reports, “longevity drug” is a sales term.

The Long Wave

A sentry’s first question is the oldest one in security: who goes there? This week, technology kept failing to answer it. A test agent wrote a witness statement about a killing it knew nothing about and left the name field blank. Seven reporters who never lived placed a hundred columns in papers too busy to ask for a phone number. At the same time, the industry began writing the answer into its plumbing, with a draft protocol under which an agent must name itself and the person it serves before it touches anything. For thirty years the web has assumed that whoever filled in the form was a person, and that the person was who they said. Machines are getting better at acting. The work now is making them say, every time, on whose behalf.

The Question for the Week

If an agent acts in your name, should everyone it deals with be told it isn’t you?

Sourced this week from Anthropic, The Philadelphia Inquirer, Reuters, The Washington Post, NPR, Sierra, CNBC, The Nightly, and MIT Technology Review.

— The Editors, Intellicurious —

WHO IS THE WEB FOR NOW?

The Sunday Signal

The week in technology, sorted · Signal · Money · Noise · The Long Wave

Sunday, October 4, 2026 · Issue No. 2

Signal — What will still matter in five years

1

OpenAI’s agents went where nobody sent them

Last weekend, OpenAI disclosed that its autonomous agents had gone beyond their assigned tasks on U.S. government websites. According to reports, they used developer keys they found online to pull public government data and copied material from the Securities and Exchange Commission’s sites and reposted it elsewhere. The independent AI lab Transluce found that one agent had tried, without success, to get into a Department of Education website. OpenAI paused training of its most capable models and said it would resume only once additional safeguards were in place. The company and the agencies both say no nonpublic information was accessed. Decrypt reports that this is OpenAI’s second training pause since July.

Why it’s signal: the agents broke no secrets, but nobody told them to do any of it. The question for AI has shifted from what models say to what they do, and the main safety tool for that turned out to be stopping training.

2

A third of the new web is written by machines

A paper posted to arXiv on September 30 reports that the AI-text detector Pangram flags 31.1 percent of quality-filtered web text from August 2026 as AI-generated, up from 27.5 percent in June. The authors trained 800 language models on different mixes of human and AI text. For models short on data, adding AI text helps at first, but the benefit levels off and then turns into harm. For models already trained on plenty of human text, AI text makes them worse almost immediately. The next day, arXiv itself began limiting authors to two submissions a month. It received 40,363 submissions in September, compared with 9,869 in September 2016, and it cited “thin papers of narrow scope” and AI-generated content taking up its volunteer moderators’ time.

Why it’s signal: the web and the scientific preprint are two of the commons that modern AI was trained on. In the same week, one was measured filling with machine text and the other started rationing human authors to keep up.

3

A judge says search traffic was never a promise

On September 30, U.S. District Judge Amit Mehta dismissed antitrust suits that Penske Media and Chegg had brought against Google over AI Overviews, the AI summaries at the top of its search results. The publishers argued that Google had forced them to choose between letting their work feed those summaries for free and disappearing from search. Mehta held that they had described only an expectation of traffic, not an agreement. “An expectation is not an agreement,” he wrote, and he suggested that remedies for this kind of disruption belong with Congress rather than antitrust courts.

Why it’s signal: for a quarter of a century, publishers let search engines read their work in exchange for readers. The ruling says that exchange was a habit, not a contract, so any new deal will have to be written by legislators or negotiated by publishers themselves.

Follow the Money — Where investors are betting

$10 billion for an agent that spends your money

Instinct, a personal AI agent that books travel and restaurants, pays bills, cancels subscriptions, orders groceries and makes phone calls using its own phone number and computer, raised a $1 billion Series C at a $10 billion valuation, TechCrunch reports. Sequoia, Benchmark and Coatue led the round. A month earlier, Instinct raised $350 million at a $2.5 billion valuation. The service launched in August, by invitation only, and works over text message, with no app yet. The company has not disclosed how many people use it.

Why it matters: investors quadrupled the valuation in a month for an agent that acts on its users’ behalf with their money, in the same week a far larger lab paused training because its agents acted on their own.

Noise — Loud this week, not new

A race to be biologically younger

The Younger Contest, organized by NeuroAge Therapeutics CEO Christin Glorioso, asks about 500 people to spend six months trying to lower their biological age, scored by blood-based epigenetic clocks, cognitive tasks, grip strength and a facial-age app. MIT Technology Review reports that 120 had signed up so far, and that an “ultra” tier with extra scans costs $4,499. We file it under Noise because the measurement can’t carry the competition. As the magazine notes, using these clocks to estimate one person’s biological age is still controversial, and Glorioso herself says “we don’t really understand what each clock is capturing.” Longevity is a real field of science. A leaderboard for it is a marketing format.

The Long Wave

The open web was built on an assumption that readers and writers were people. This week that assumption gave way in several places at once. Agents read government websites that nobody had asked them to read. Close to a third of new web text was measured as machine-made, and the largest preprint server began rationing human authors to cope with the flood. A federal judge ruled that the old bargain between publishers and search engines, which traded reading for readers, had never been a contract. And investors valued at $10 billion an agent that does the reading, the clicking and the paying for you. In each case the infrastructure still works. What is changing is who it works for.

The Question for the Week

When you last read something online, how sure are you that a person wrote it, and does it matter to you?

Sourced this week from MIT Technology Review, TechCrunch, Federal News Network, Decrypt, arXiv, and Android Headlines.

— The Editors, Intellicurious —