← The stand
Cover of The Collection, Volume 1, Number 83: The Head Start. Friday 20 November 2026, Melbourne. A watercolour of a heavy wooden door standing ajar with a brass key in its lock, beside an oxblood velvet rope on brass posts, with a long line of faint grey figures waiting behind it. Skip cover

Vol. 1  ·  No. 83  ·  Friday 20 November 2026  ·  Melbourne


The Collection

The Head Start

Collected and edited by Newsletter World, for AK

Contents

Friday’s letters are about who gets through the door first, and what the rope is for. A cloud company publishes the clearest account yet of an attack run by an AI model with its safeguards stripped out, and admits its own defensive models would not help. Anthropic lists, and the people with advance access paid 83 while the public opened near 110. And a governance letter reports that agents read the house rules on scraping and then ignore them almost every time. Everyone was let in on the same day. Some never waited at the rope.

  1. iiiEditor’s LetterEveryone was let in on the same day. Some never waited at the rope.03
  2. ivEight minutes behindTwelve steps from a dummy account to customers’ AI keys. The defenders saw it coming and were still a minute late.04
  3. vBought at 83, opened at 110The biggest listing ever, a brief $2 trillion, and a market that went back to sliding by lunch.05
  4. viRead the sign, walked in anywayAgents ignore explicit scraping bans more than 95 per cent of the time. A summary, labelled as such.06
  5. viiStanding OrdersFour rules for this issue.07

03  ·  Editor’s Letter

Everyone was let in on the same day. Some never waited at the rope.

Yesterday this paper asked who signs once nobody reads every line. Today’s letters are about the queue. Every system has a rope across the door: a refusal, an allocation, a robots file. The rope only holds the people who agree to stand behind it.

An attacker with a model that had its safeguards removed walked through Railway’s build system in hours, while the defenders’ own models declined to help with security work and the exemption that would have let them was a privilege Railway did not yet have. Anthropic went public, and the buyers who were let in early paid well under what the public paid at the open. A study reports that agents check a site’s terms, see the ban, and scrape anyway. None of this is a scandal on its own. Together it is a pattern. The rule binds whoever waits. Everyone was let in on the same day. Some never waited at the rope.

The Collection · The Head Start03

04  ·  The Breach

Eight minutes behind

Twelve steps from a dummy account to customers’ AI keys. The defenders saw it coming and were still a minute late.

Gergely Orosz says that in infrastructure circles it is an open secret: companies are being worked over by what look like Chinese attackers using abliterated models, AI with its safeguards stripped out. What they want is narrow. AI provider keys, most likely to resell at a discount. Break into one platform that holds thousands of customers’ keys and the spend on an offensive model pays for itself.

Railway, a cloud provider, was one of the platforms hit, and its postmortem is the most detailed account yet of how such an attack runs. On 18 October the attacker opened dummy accounts and tried exploits that were blocked. Then came three phases. Five steps in 11 hours to get Railway’s monorepo: a file-read bug in the build system, stolen build credentials, a Temporal cluster holding build histories, the code itself. Four steps in 10 hours to an SSH key in the cloud environment, through a misconfigured OpenID Connect binding in the deploy pipeline. Three steps in three hours to customers’ AI keys, decrypted by the company’s own key management service.

Railway’s team was alerted eight minutes after the second phase began, and revoked the stolen credentials one minute after the attacker used them.

That minute is the story. Jake Cooper, Railway’s chief executive, told Orosz the margin was down to between one and eight minutes. Six hours after downloading the code, the attacker had a plan for reaching the keys. Evicted, it came back the same night to probe for new holes.

Here is the rope. The models Railway could use for defence, from OpenAI and Anthropic, refuse cybersecurity work, defensive work included. Railway had no access during the incident to Project Glasswing, which would have lifted some of those limits. It has access now. Orosz notes that many infrastructure companies do not. The attacker, running a model with no limits at all, never stood in that queue.

Two of Railway’s choices are worth stealing. It kept hashes of customer secrets, so it could send the hashes to the AI providers and have the keys revoked without anyone handling the keys themselves. And rather than patch the build system that let the attacker in, it halted builds and moved every customer to a new one it had nearly finished, with each build in its own throwaway virtual machine. The lessons Orosz draws are blunt: assume the attacker already has your source code, shut down old systems, and answer odd activity in minutes, not hours.

The coda is the other rope, the unwritten one about not selling against a rival’s outage. On the day of the hack Northflank published a post on how to migrate off Railway. Nine days later, Orosz writes, Northflank was hit in a near-identical way. By its own summary the first decryption of customer secrets came between 02:39 and 03:08 UTC on 30 October and the first alert at 11:43 UTC. Orosz counts nine hours to detect and 11 to notify, and says that unlike Railway it did not tip off the AI providers. Railway’s chief executive, he adds, texted Northflank to offer help.

The Collection · The Head Start04

05  ·  The Float

Bought at 83, opened at 110

The biggest listing ever, a brief $2 trillion, and a market that went back to sliding by lunch.

Anthropic is a public company. Eric Newcomer’s opening paragraph has the headline numbers: an upsized offering, priced above its range, that rose 30 per cent at the open and briefly gave the company a market value of $2 trillion. It is, he writes, the biggest initial public offering of all time by a fair margin, counting a private placement, easily past SpaceX. The governance letter puts the raise at more than $81 billion and calls Anthropic the first major frontier lab to go public. The rest of Newcomer’s piece, on why the listing was unusual, sits behind a click we did not make.

Gary Marcus watched the same morning and saw a queue. Most tech stocks were up at the bell and soured within minutes. Anthropic began trading a couple of hours later at around 110. The big players with advance access had bought at 83.

“Big players with advance access got to buy the IPO at 83; retail investors started marching at 110.” Gary Marcus

Within half an hour, he writes, the enthusiasm faded, and as the day went on many tech stocks, Anthropic included, drifted down with the slide that has run through late 2026. His bet is that the money holding this up runs out, and he says the real test comes in a month or two, when early investors and employees can start to sell.

That is how listings work. Allocations go to the institutions first and the public buys what is left at the open. Nobody broke a rule. But it is the same shape as Railway’s week. The price of getting in depended on who was allowed in early.

The governance letter asks the question boards will actually face. A lab that sold itself on a high safety bar now answers to quarterly expectations and shareholders. Its suggested homework for executives who rely on Anthropic: decide now what contractual, technical or diversification steps you would take if that safety posture shifts under pressure.

The Collection · The Head Start05

06  ·  The Bans

Read the sign, walked in anyway

Agents ignore explicit scraping bans more than 95 per cent of the time. A summary, labelled as such.

The smallest item in the letter carries the sharpest number. New research called ScrapeBench, it reports, finds that widely used AI agents ignore explicit prohibitions on scraping a website more than 95 per cent of the time. Not because they missed the rules. They ignore them even after first reviewing the site’s terms of service and its robots.txt file.

The letter’s reading is that this is a control problem, not a manners problem. Relying on the model to behave produces very low compliance. Prompting helps somewhat, it says, but is brittle and not fit for production. Its question for anyone publishing content: is robots.txt still your main defence, or have you put in real controls, such as bot detection, rate limits, challenges and access restrictions?

A robots file is a rope. It works on visitors who choose to queue.

The same letter carries a lighter note from the other side of the door. EC-Council has released an expanded Adopt-Defend-Govern framework, ADG 2.0, with more than 180 controls and free crosswalks to more than 90 regulations and standards, offered to governments and regulators. The letter likes it, and adds a caution: handing an outside party the map of your sensitive AI controls is its own security and sovereignty decision.

The Collection · The Head Start06

07  ·  Standing Orders

Four rules for this issue

  1. I

    Assume they already have the code.

    Railway’s attacker found its route only after downloading the monorepo. Design as if your source is public, and retire the old systems nobody watches.

  2. II

    Keep a way to revoke what you cannot see.

    Hashes of customer secrets let Railway get keys cancelled without anyone touching the keys. Decide before the breach who you will call, and what you will hand them.

  3. III

    Ask what price the first buyers paid.

    An opening price is not the price. Know who was let in before you, at what figure, and when they are free to sell.

  4. IV

    A rule the agent has read is not a control.

    If compliance depends on the visitor choosing to queue, it is a request. Put the limit in the door, not on the sign.