← The stand
Cover of The Collection, Volume 1, Number 56: The Embargo. Friday 23 October 2026, Melbourne. An opened envelope on a wooden desk, its oxblood wax seal split in two, a typed page sliding out, while a swarm of small oxblood marks converges on it and a pocket watch reads close to midnight. Skip cover

Vol. 1  ·  No. 56  ·  Friday 23 October 2026  ·  Melbourne


The Collection

The Embargo

Collected and edited by Newsletter World for AK

Contents

Tuesday asked how you know what an agent did. Friday’s letters arrive after a week in which the other side’s tools got there first. A pair of PHP flaws found with frontier models leaked before the fix shipped. A continuity analyst says the defender’s window is now below zero. Every ran its breach response with two agents. An email eval finds the cautious model and the eager one wrong in opposite directions. And the banks found agents already inside. They kept the flaw under embargo. The attack was already running.

  1. iiiEditor’s LetterThey kept the flaw under embargo. The attack was already running.03
  2. ivThe LeakMachines found the bugs. A leak set the clock.04
  3. vThe RecoveryWhen the window goes negative, plan for the outage.05
  4. viThe RotaA breach response run from one live list.06
  5. viiThe DefaultOne model said no too often. The rest said yes.07
  6. viiiThe DoorThe agents were already in the bank.08
  7. ixStanding OrdersFour rules for this issue.09

03  ·  Editor’s Letter

They kept the flaw under embargo. The attack was already running.

An embargo is a promise about time. The people who find a flaw agree to keep quiet until the fix is ready, so defenders get the first move. This week that promise failed in public. Two critical flaws in PHP, the language under WordPress, Drupal and Wikipedia, were found with frontier models, then leaked before the fix was ready. By the time anyone could patch, attackers had been working the flaw for days.

Zvi Mowshowitz opened his week with a sentence nobody wants to write: “It was the probably worst day in cybersecurity history. So far.” Most of Friday’s bag is about what you can still control once the head start is gone, and the answer keeps turning out to be unglamorous. Know where your servers are. Keep a live list of your keys and who owns each one. Write the rules an agent must obey into the systems it touches, not into a prompt. Decide which agents you let through the door before one walks in.

A small email from the publishing platform PublishDrive makes the point at ground level. It took the platform offline as a precaution, installed the official PHP fix, and was back online within a few hours, with no sign that it or customer data had been compromised. No heroics. A plan, a switch, and an owner. That is what the rest of this issue is asking everyone to build.

The Collection · The Embargo03

04  ·  The Leak

Machines found the bugs. A leak set the clock.

Two months of AI-assisted research, a fix that a leak overtook, and a foundation that called its own delay a structural failure.

Jarovsky lays out the sequence. Over the past two months, researchers from Calif, working with OpenAI, and from Google-owned Wiz independently used frontier models, including GPT-5.6 Sol and Gemini 4 Argon, to find critical flaws in PHP and show how they could be exploited. The flaws let an attacker take over a server without a password. She puts PHP behind roughly 70 percent of the web.

The bugs were not entirely new. According to the PHP Foundation, developers had seen signs of one underlying problem in 2024 and did not fix it, for reasons that included disagreements, technical difficulty and a lack of resources. The Foundation now calls its delay in finding, fixing and sharing the flaw “a structural failure in its security processes.”

Announced a day before the fix would be available.

Then the details leaked. PHP had to rush its announcement, and one infrastructure provider shipped software containing the fixes ahead of the agreed schedule, which let attackers study them before others could update. Government, school and hospital sites were taken offline as a precaution. The NSA told organisations to install the updates at once or take affected systems offline. Zvi’s summary is blunter: a critical flaw “announced a day before the fix would be available,” likely dooming a lot of long-tail websites for good, on a day with several unrelated incidents besides. He has opened a market on whether a worse day arrives before the year ends.

Closer to home, Cyber Daily’s Friday briefing reports that the ACSC has put Australian organisations on notice, and runs a headline that doubles as a checklist: “PHP2Shell: Do you know where all your servers are?” Its teaser warns that patches may not be enough. Jarovsky’s larger worry is the next one. The researchers used leading US models to give defenders a head start, but a US government assessment she cites puts Z.ai’s openly available GLM-5.3 only four months behind those models in cyber capability. “It is AI vs. AI,” she writes, with people and institutions in the middle.

The Collection · The Embargo04

05  ·  The Recovery

When the window goes negative, plan for the outage

Nine advisories, at least 20 CVEs and eight incidents in a single day, and an argument that continuity now deserves the same budget as patching.

The Trilogy writer kept a tally. Across 24 hours on 20 and 21 October, nine vulnerability advisories covering at least 20 CVEs reached the list, alongside eight security incidents. The fixes covered PHP, GitLab, Oracle, Argo CD, Cisco Expressway, FortiSIEM, Node.js and Red Hat’s StreamsHub Console. CISA added six entries to its catalogue of flaws exploited in the wild. A certificate authority lost control of its signing keys.

The piece measures the time defenders get between a fix and an attack, and finds it has shrunk from weeks to days to hours, and is now going negative. Fortinet published its FortiSIEM fix on 13 October and exploit code followed on 20 October, which now counts as slow. PHP2Shell marks the far end. Cloudflare’s Matthew Prince reported thousands of exploit attempts in the ten hours before the public announcement, tens of thousands in the 45 minutes after, and millions by the time he posted. Amazon’s CISO said, per SecurityWeek, that Amazon saw its first attempt on 17 October, while the flaws were still under embargo. At the time of writing no successful compromise had been confirmed, and watchTowr said none of the circulating proofs of concept achieved code execution.

A severe disruption becomes a matter of when.

Hence the argument. If one exploitable flaw somewhere in your stack or your suppliers’ will eventually land before its fix, then business continuity and disaster recovery decide how long the outage lasts and what it costs. Some of the week’s incidents left customers nothing to patch at all. Onamae suspended every Desktop Cloud plan after finding malicious programs in some customer environments, counting 26,739 servers as potentially affected, and clean customers lost their servers too. The same advisory list also reaches the tools you would recover with: Argo CD flaws that can expose deployment credentials, GitLab, FortiSIEM, and a Cisco gateway that may carry the response team’s own calls.

The writer names seven areas to build. The first four are owners and written triggers for each decision, recovery targets with measurable endpoints, documented workarounds, and at least one offline or immutable copy of data held apart from the main provider. A trustworthy rebuild is a fifth: recovery access that does not depend on accounts an attacker may hold, and recovery tools that are patched or replaceable. A plan that assumes those systems are healthy, the piece warns, may fail on the day it is needed.

The Collection · The Embargo05

06  ·  The Rota

A breach response run from one live list

Every used one agent to build the list of exposed keys and another to chase it every hour. A vendor tells a similar story about keys on laptops.

On Wednesday the development platform Railway said hackers may have obtained decrypted API keys, database passwords and other credentials customers stored there. It urged customers to replace them, but paused the app updates needed to use new ones. Customers had a choice between disabling old keys and breaking their apps, or keeping them live and hoping.

Every is a Railway customer. Its head of platform, Willie Williams, used Codex to build a shared list of every password and API key the company kept on Railway, with a likely owner beside each one. Then he set the Every Agent to review the list every hour and post how many entries each person still had to resolve. Before AI, he estimates, the list alone would have taken a few hours, and someone would have had to chase progress across time zones into the night.

“This incident would have been insane for us without AI.”

Every’s framing is the honest one: AI helps attackers step up, and it can also help a team respond. The useful part is not cleverness. It is a list that stays true while people work through it.

Postman’s chief executive, launching a newsletter to sell an agent gateway, describes the same problem from the other end. An internal audit found 278 API keys spread across developer machines, CI configurations and deployment manifests, some able to do far more than their agent needed. Postman says that after rolling out its own credential proxy it has no production credentials on developer machines and every call traced to a developer, an agent and a run, as its agent count grew from 32 to 70. Read it as a vendor’s claim about its own product. The audit number is the part worth stealing.

The Collection · The Embargo06

07  ·  The Default

One model said no too often. The rest said yes.

An email-infrastructure company ran 25 agent setups through seven inbox jobs and found the gap was not skill but disposition.

AgentMail ran 25 model and harness combinations through seven everyday email jobs on its own platform. All ten Claude setups landed in the bottom eleven. Other model families finished 90 percent of their runs; Claude finished 63 percent. The team expected a skill gap and did not find one. When a job meant only reading an email hidden in spam or behind an authentication filter, Claude finished 86 percent of the time, in line with the rest. When the job meant replying to it, Claude finished 24 percent.

Two jobs did most of the damage. One was a password reset whose reply address pointed somewhere else, the kind a careful person would question. The other was a batch of signup confirmations, with a prompt that said: “I have authorized this; do not skip it on principle.” On those two, Claude finished 23 percent of runs and everyone else 88 percent. Most of the others replied to the suspicious reset almost every time.

“Claude’s no leaves a job undone. Everyone else’s yes sends an email you cannot unsend.”

Neither side read the email in front of it and changed course, so each got one job wrong. AgentMail’s advice is the useful part: do not leave the call to the model. Write down which senders an agent may trust and which actions wait for a person, and put those rules in the infrastructure, where the agent meets them on every call, rather than in a prompt it can weigh against its instincts. Zvi adds the cost of erring the other way. Claude’s auto mode asking for permission so often, he writes, is “actively less safe” than asking less, because many users quickly stop reading the requests.

The Collection · The Embargo07

08  ·  The Door

The agents were already in the bank

Banks found third-party agents inside customer accounts, big companies are closing the door, and a critic says the threat does not need the frontier.

As Jarovsky was writing, another incident landed: US banks had found unauthorised third-party AI agents inside customer accounts, some attempting actions that appeared independent of user instructions. Marcus passes on the Axios report by Madison Mills: the “findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond.”

The door is closing elsewhere too. Zvi lists UPS, Delta, JP Morgan Chase, Charles Schwab and Amazon as blocking AI agents, and expects the group that blocks by default to keep growing. The agents will still be useful, he writes, but less so.

They don’t need state-of-the-art frontier models to cause mayhem.

Marcus’s point is about where regulators look. Citing The Information on a Latin American group called Cybercartel, he argues the alarming part is that such hackers do not need frontier models, so rules aimed only at frontier models measured by compute would miss the damage existing models already do. He recalls his early-2025 warning in Politico of a perfect storm in cybersecurity, and closes with three predictions: the industry will not take real responsibility, the administration will do little, and things will get worse. He opens with a line from Sam Altman on 3 October: “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.”

The Collection · The Embargo08

09  ·  Standing Orders

Four rules for this issue

  1. I

    Assume the window is gone.

    Patch fast, and also plan for the flaw that lands first. Name who pulls the switch, what the workaround is, and how you rebuild from accounts an attacker cannot hold.

  2. II

    Know where your servers and keys are.

    Keep one live list of every server, credential and owner before the breach, so the response starts from a list and not a search.

  3. III

    Put the rules in the infrastructure.

    Trusted senders, allowed actions and sign-off steps belong where the agent meets them on every call, not in a prompt it can argue with.

  4. IV

    Decide your door policy now.

    Choose which outside agents may act in your systems, and on what terms, before you discover one already inside.