Day Lila

Cyber · Thursday, 19 November 2026

Written by AI agents built on Claude, from the sources they cite, and overseen by Zeemish. How it's made

This day: every beat The day in 12 minutes

Swipe 28 stories

Five teams find the same critical flaws in NGINX, the web server behind a third of websites, and F5 ships fixes

28 stories 47 sources

Signals on this beat

Fixes arriving in floods High ↑
  • Oracle released fixes for about 948 flaws this month. [37]
  • Google's Chrome 158 fixed 204 more. [38]
  • The team behind PHP set a critical update for Thursday. [39]
Patching that breaks things Building ↑
  • US officials told Congress that patching and upgrades, not hackers, lay behind last week's outages. [10]
  • A source said the NSA's network faltered during maintenance on PHP, a common programming language. [11]
Old flaws newly attacked High ↑
  • Attackers are using a flaw in Apache Roller, a blogging server, that reaches back to version 4.0.0. [33]
  • A flaw in Puppet, a tool that manages company computers, joined CISA's list of attacked flaws on 17 November. [34]
Trusted names used as bait High ↑
  • Russian spies posed as a real employee of a firm they had hacked. [21]
  • Scammers used an AI-made Elon Musk to ask viewers for crypto. [29]
  • Attackers put malware out under Microsoft's own package name. [20]
Suppliers' breaches reaching customers Building →
  • A break-in at Wellington IT, a Belfast company, exposed data from about 40 credit unions. [15]
  • Flaws at CreditNirvana, an Indian debt collector, exposed customers of three big Indian finance firms. [30]

The stories

28 on this beat today.

Each story is told in its parts: what happened, the questions that explain it, how it works, who gains and who pays, and the good news. A part appears when the sources give it.

  1. 01

    Five teams find the same critical flaws in NGINX, the web server behind a third of websites, and F5 ships fixes

    225 million

    websites run on NGINX, about a third of the web

    F5 maintains the free, open-source version [3]

    9.2 of 10

    danger score of each of the 5 flaws

    the worst needs no login to run code [2]

    98.9%

    of 11.94 million checked NGINX servers on affected versions

    from Xint's snapshot of servers whose version it could read [2]

    8 times

    more flaws found by leading tech firms since April

    after Anthropic released its Mythos model [5]

    What happened

    • NGINX is software that runs about 225 million websites, around a third of the web. [3]
    • Five research teams, working separately, found 5 critical flaws in it this week. [4]
    • The worst lets an attacker run their own code on a server from anywhere, with no login. [2]
    • Each of the 5 flaws scored 9.2 out of 10 on the standard danger scale. [4]
    • The worst sits in HTTP/2, the part that lets a browser send many requests down one connection. [1]
    • Xint, an AI security firm, joined one flaw to a separate Linux bug and took full control of the machine. [2]
    • Wiz, Ikotas Labs, Metnir, Rapid7 and Xint each found flaws while working independently. [3]
    • Most of them, if not all, used AI to help, The Register reports. [3]
    • Xint pointed its tool at about 50,000 lines of NGINX code. [2] 90 minutes later it reported serious flaws. [2]
    • Xint says its tool found 4 of the 5 flaws with no person guiding it. [2]
    • Since April, the flaws found by leading tech firms have grown 8 times over, the Guardian reports. [5]
    • That rise followed Anthropic's release of Mythos, an AI model with a startling ability to flag flaws. [5]

    Who is involved

    F5, the company that looks after NGINX, has released fixed versions. [3]

    • F5
    • Wiz a security firm owned by Google Cloud; it found the two HTTP/2 flaws
    • Xint an AI security firm; its tool found four of the five flaws alone
    • Rapid7 a security firm; it built a working test attack

    How it unfolded

    The worst flaw's code goes back to version 1.9.5, released in 2015. [3]

    It follows PHP2Shell last month and Java4Shell a day earlier, both also found with AI. [3]

    1. 2015 the flawed HTTP/2 code ships in NGINX 1.9.5 [3]
    2. This week five teams report NGINX flaws; F5 ships fixes [4]
    3. October PHP2Shell
    4. A day earlier Java4Shell, another AI-found flaw, is disclosed [2]
    5. This week NGINX
    6. April Mythos is released; flaw finds grow 8 times over [5]

    Why it matters

    Security teams must now find every NGINX copy, test the fix and roll it out, Wiz says. [3]

    Switching HTTP/2 off is the stopgap, but services that need it, like Google's gRPC, stop working. [1]

    Think it through

    1. 1

      Why did five teams find the same bug at once?

      Wiz linked it to AI getting better at tracing memory mistakes in code. [2]

    2. 2

      Can attackers use those tools too?

      That ability is open to defenders and attackers alike, CyberScoop notes. [2] Forescout, a security firm, says the same of its research. [6]

    3. 3

      How much faster is the hunting now?

      Forescout says AI cut one research job from about 9 months to 90 hours. [6]

    4. 4

      So are websites being hacked through it now?

      F5 says it has found no evidence of attacks so far. [3] Xint says the bugs are old enough that past attacks cannot be ruled out. [4]

    5. 5

      How many servers still run the flawed code?

      Xint counted 98.9% of 11.94 million NGINX servers it could check running affected versions. [2]

      98.9%of checked NGINX servers on affected versions

    6. 6

      What does this have to do with the outages?

      US lawmakers say the race to patch AI-found flaws helped cause last week's outages. [9]

    7. 7

      So who decides how fast this hole closes?

      F5 has shipped the fix, so each website's owner now decides when it goes in. [3] Rapid7's public test attack needs about 128 GB of traffic, a real limit. [2]

    How it works

    Software engineers call these memory-safety bugs. [3] The code writes data outside the memory set aside for it. [1]

    NGINX kept a running count of how much data the other side may receive. [1] An attacker pushes that count past its limit, so the memory-safety bug spills data out. [1]

    Seen it before

    Pour from a jug sized by a wrong count and the glass overflows onto the table.

    The truth

    • What is really going on AI tools are now finding old flaws in software nearly everyone runs, NGINX this week after Java and PHP, and US officials say the rush to patch them helped cause last week's outages. [3][9]
    • Xint's 98.9% counts software versions, not servers truly open to attack. [2] Some may have HTTP/2 off, or a fix copied into an older version. [2]
    • The teams count the flaws differently: Ikotas Labs reports 1, Wiz 2 and Xint 5. [4]
    • Who gains AI security firms — Each big disclosure showcases their tools: Xint's chief called the NGINX find a testament to its AI harnesses. [2]
    • Who pays IT and security staff — They must find, test and install each round of fixes, often all at once, on systems not built for that pace. [3][9]
    • Nobody knows yet

      When the Linux flaw that Xint joined to NGINX will be fixed.

      It had no fix and no public ID when NGINX's flaws were disclosed, and Xint is holding back its details until there is one. [1]

    Good news

    • F5 shipped fixes before any attack using the flaws was seen. [3]

    What comes next

    F5 says it is still checking whether its own cloud services are affected. [2]

  1. 02

    US blames patch rush for outages

    What happened

    • US officials told Congress on Wednesday that hackers did not cause last week's string of outages. [10]
    • The outages hit Microsoft's email, US flights, big banks, Apple Pay and the NSA, the US spy agency. [5]
    • Microsoft blamed a network upgrade, the FAA a router upgrade, and Chase a software bug. [5]
    • Jay Clayton, the US intelligence chief, said "the rate of patching" is increasing. [9]
    • A source told The Record that the NSA outage came from maintenance on PHP, a common programming language. [11]

    Who is involved

    • Microsoft email down
    • US flights grounded
    • Banks accounts hard to reach
    • NSA network offline

    Why it matters

    Senator Mark Warner says the rush to patch everything at once strains systems not built for it. [9]

    Think it through

    1. 1

      Why is everyone patching at once?

      Hospitals raced to "patch before they hack", their US association says. [8] They assume attackers are using AI to find the same flaws. [8]

    2. 2

      Are attackers using the chaos to break in?

      The intelligence, homeland security and FBI officials said they see no sign of that so far. [9]

    3. 3

      So is the fast patching worth the outages?

      Representative Jim Himes says the fast pace is "vital to shore up the security" of systems. [9] He says faster patching could cause more frequent outages. [9]

    How it works

    Engineers call it technical debt: quick fixes piled up instead of solving the real problem. [7]

    The US flight stop came while engineers swapped old copper wires for fibre. [7] A build-up of quick fixes can cause more serious problems later, the BBC explains. [7]

    Seen it before

    A cupboard stuffed for years spills everything the day you finally open it to tidy.

    Why a hidden flaw one team finds is soon found by others

    Vulnerability researchers call it rediscovery: when many hunt with similar tools, a flaw one team finds is likely found by others, attackers included.

    The twist

    When five friendly teams find the same flaw in the same week, nobody can safely assume an unfriendly team has not.

    The picture

    Rediscovery is strong when hunters use similar tools on the same code. On the left, five teams converged; on the right, three different AI models mostly found different things.

    Step by step

    1. A flaw sits hidden in widely used code for years
    2. Better search tools reach many hunters at about the same time
    3. Several teams find the same flaw on their own
    4. So anyone else hunting, attackers included, may have it too
    5. Defenders treat a flaw one team found as already known

    The same force, elsewhere today

    Where this chain is also running, in today's other stories.

    • Microsoft's AI agent kit poisoned

      several security firms found and analysed the same poisoned upload independently, because they all watch the same public package store

    • US blames patch rush for outages

      hospitals raced to patch Java because they assume attackers with AI tools are hunting the same flaws defenders just found

    Seen it before

    Science

    calculus and the theory of evolution were each worked out by two people at about the same time

    Patents

    rival inventors often file for the same idea within days, because they started from the same new tools

    Exam leaks

    once one copy of the answers is out, the school assumes others have it and sets a new paper

    The catch

    It holds only when hunters search alike: in Echo's test, three different AI models agreed on just 1 of 121 possible flaws.

    And the whole of it

    The teams that found NGINX's flaws and anyone hunting them for harm use the same kind of tool. The owners of 225 million websites sit between them, mostly unaware either is looking.

    The truth

    • Why it works on people Several failures in a few days feel like one attack: Senator Chuck Schumer called it "too coincidental", though each was traced to its own upgrade or bug. [11][5]
    • The officials blamed patching but did not say the patches were for AI-found flaws. [9] Lawmakers made that link. [9]
    • Nobody knows yet

      Whether the NSA's outage was patching an AI-found flaw.

      A source confirmed PHP maintenance was the cause, but said it is unknown whether that flaw was found by AI or by people. [11]

  1. 03

    Equinor fights an attack on its systems

    What happened

    • Equinor, a Norwegian energy company, is working to contain an attack on its computer systems. [12]
    • Its oil and gas output has not been disrupted, and it sees no sign that data was stolen. [12]
    • Staff may have trouble reaching company apps, and its internal email was down, Reuters reported. [12]
    • Equinor said "It is too early to determine how long this will take." [12]

    Who is involved

    It is state-owned, and one of the largest oil and gas producers in Europe. [13]

    • Equinor a big gas supplier to Europe

    Why it matters

    The attack comes amid wide concern about the safety of Europe's vital services. [13]

    Think it through

    1. 1

      Could the gas stop flowing?

      An industrial security expert says the question is how long output lasts without the business systems behind it. [13]

    2. 2

      So who decides when it ends?

      Equinor is working with government security agencies to clear the attack. [13] It has given no date for full recovery. [13]

    How it works

    Incident responders call the first step containment: cut off what the attacker can reach, then repair. [12]

    Equinor is still in containment, working to "neutralise the attack" before restoring its systems. [12]

    Seen it before

    A plumber shuts off the water before mending the burst pipe.

    The truth

    • Equinor says no virus was involved, so nothing can spread. [12] It has not said who attacked or how they got in. [12]
    • Nobody knows yet

      Who attacked Equinor, and how they got in.

      Equinor has not named the attackers, said how they got access or which parts of its systems were hit. [12]

  2. 04

    Breach at IT firm hits 40 credit unions

    What happened

    • A break-in at Wellington IT, a Belfast company, exposed data from about 40 credit unions in Ireland. [15]
    • An automated program made 5 million requests to use a flaw in how Wellington checked access. [14]
    • Exposed were the names, bank account numbers and payment references of people that members pay. [14]
    • The attack ran from 2 to 4 October. [14] The credit unions were told only this Tuesday. [15]

    Who is involved

    Credit unions are non-profit banks owned by their members. [15]

    How it unfolded

    1. 2-4 October the attack
    2. Tuesday credit unions told

    Why it matters

    Members' saved payees could help scammers make fake payment requests look real. [14]

    Think it through

    1. 1

      Why were members told so late?

      Wellington IT took weeks to tell the credit unions, a delay junior finance minister Robert Troy called "not acceptable". [15]

    2. 2

      Does anyone know whose data was taken?

      No: Wellington IT says it lacks the logs to show which members' data was reached. [14]

    3. 3

      So who has to tell the members?

      The credit unions must, Troy says. [15] But they depend on Wellington IT for the facts. [15]

    How it works

    Security managers call this third-party risk: your safety rests on the suppliers holding your data.

    That third-party risk is plain here: about 50 credit unions use Wellington IT, and members of about 40 were exposed. [14]

    Seen it before

    Give one dog-walker the key, and every house on the walker's round shares one risk.

    The truth

    • What is really going on The week's break-ins went through trust instead: a poisoned Microsoft package, Russian spies posing as a real employee, and a supplier's flaw that reached 40 credit unions. [18][21][15]
    • No member has lost money so far, Troy says. [15] The full extent is still being established. [15]
  3. 05

    London travel card data accessed

    What happened

    • Hackers reached the details of about 265,000 Oyster photocard holders, Transport for London (TfL) says. [16]
    • Names, home addresses, birth dates and security answers were reached, but no passwords or bank details. [16]
    • TfL took the photocard website offline and told the UK's data regulator. [16]

    How it unfolded

    It is the second cyber incident to shut the system in just over two years. [16]

    1. September 2024 first attack
    2. Monday suspicious activity found

    Why it matters

    Until the site returns, nobody can apply for or renew a photocard online. [16]

    Think it through

    1. 1

      How bad was the last breach?

      In 2024 TfL said about 5,000 people were contacted; the BBC later found about 10 million people's data was stolen. [16]

    2. 2

      So who checks TfL this time?

      The Information Commissioner's Office, the UK data regulator, which cleared TfL over the last breach. [16]

    How it works

    Data protection law calls this a personal data breach: someone reached personal details they should not have.

    TfL reported this personal data breach to the ICO and took the site offline to limit access. [16]

    Seen it before

    Lose a class register and the school must tell every family on it.

    The truth

    • TfL has given no cause yet, only that "suspicious activity" was found on Monday. [16]
  4. 06

    Doctors' regulator leaks complaint data

    What happened

    • The General Medical Council (GMC), which regulates UK doctors, sent a newspaper data on 223 doctors by mistake. [17]
    • The data summarised complaints alleging antisemitism or Islamophobia, made from October 2023 to September 2026. [17]
    • Names were left out, but each doctor's GMC number was in it. [17] That number finds them on the public register. [17]
    • Some doctors first learned a complaint existed from the GMC's breach warning. [17]

    How it unfolded

    1. October 2023 first complaints
    2. September 2026 latest complaints

    Why it matters

    The doctors can now be identified over complaints the GMC is still bound to investigate. [17]

    Think it through

    1. 1

      Will the newspaper publish it?

      The Times promised not to make the details public, but said it will not delete them. [17]

    2. 2

      So who can undo it?

      Nobody fully can: the Times keeps the copy. [17] The GMC has reported itself to the ICO, the data regulator. [17]

    How it works

    Data protection experts call data with names removed but a link back pseudonymised: it can still point to a person.

    Here the GMC numbers were the link back, so the pseudonymised data named each doctor in one look-up. [17]

    Seen it before

    A class list with names blanked but seat numbers kept still shows who sat where.

    The truth

    • Some doctors say their ethnicity was in the data too; that is their concern, not a confirmed fact. [17]
  5. 07

    Microsoft's AI agent kit poisoned

    What happened

    • Attackers published 2 poisoned versions of Microsoft's Agent Framework, a tool for building AI agents. [18]
    • It is downloaded about 200,000 times a month from PyPI, the site that hosts Python packages. [18]
    • The bad versions stole passwords and keys for cloud services such as Azure, AWS and GitHub. [18]
    • They ran whenever Python started, even if a program never used the package. [19]
    • PyPI blocked the first version, and the attacker uploaded a second about 2.5 minutes later. [18]

    Why it matters

    Developers who installed either version may have handed over keys to their cloud and build systems. [19]

    Think it through

    1. 1

      How did the attacker get Microsoft's publishing rights?

      Nobody knows yet. [19] The uploads skipped Microsoft's normal release process, Wiz found. [19]

    2. 2

      Who spotted it?

      Several security firms found and documented it, including Socket, StepSecurity and OX Security. [18]

    3. 3

      So who carries the risk now?

      Every team that installed it, which researchers advise to change its keys and rebuild. [19]

    How it works

    Security engineers call this a supply chain attack: poison a trusted part, and everyone using it is hit. [19]

    This supply chain attack hid well because the poisoned package still ran the real framework underneath. [18]

    Seen it before

    Juice tampered with at the factory reaches every shop that sells it.

    The idea, in story 02: Why a hidden flaw one team finds is soon found by others →

    The truth

    • SafeDep found an account, created on the day of the attack, with rights to publish it. [19] Its role is not confirmed. [19]
    • Reports differ on timing: BleepingComputer puts the second upload 2.5 minutes after the block. [18] SecurityWeek says the two versions came 33 minutes apart. [19]

    Good news

    • PyPI quarantined the second poisoned Microsoft package about 8 minutes after it appeared.
  6. 08

    Russian spies pose as a hacked firm's staff

    What happened

    • Google says a Russian spy group used data stolen from one US organisation to hack another. [22]
    • In late 2025 it took over a top administrator account and copied the whole cloud drive. [22]
    • It stayed inside for several months, and got back in after passwords were changed. [22]
    • Later it posed as an employee of that firm to approach someone at a second US organisation. [22]

    Who is involved

    Google calls the group Cold Relic; others call it Star Blizzard. [21] US authorities link it to the FSB, Russia's security service. [22]

    • Cold Relic Russian spy group, linked to the FSB

    How it unfolded

    1. Late 2025 admin account taken
    2. Then whole drive copied

    Why it matters

    Colleagues and partners of the first victim became targets of messages from a fake insider. [21]

    Think it through

    1. 1

      How did the fake colleague seem real?

      The links led to documents about their actual conversation, which made the trap hard to spot. [22]

    2. 2

      Why steal staff lists and emails?

      The group uses stolen emails and data to build believable stories for later targets, Google says. [47]

    3. 3

      So who is the real target?

      Google says Russian intelligence turns ordinary business messages into attacks on trusted partners. [47]

    How it works

    Security researchers call this social engineering: fooling a person rather than breaking a machine. [47]

    Cold Relic's social engineering used a real colleague's identity, the thing its target trusted. [21]

    Seen it before

    A stranger who knows your teacher's name gets past the school gate more easily.

    The truth

    • Google has not named either US organisation or said what was taken. [22]
  7. 09

    UAE details Iran's cyber attacks

    What happened

    • The UAE says cyber attacks reached up to 800,000 a day at the conflict's height, mostly driven by Iran. [23]
    • Iran used wipers, programs that erase data instead of holding it for ransom, he said. [23]
    • The targets included power, water, desalination, school and hospital systems. [23]

    Who is involved

    Mohamed Al Kuwaiti, the UAE's cyber security head, set out the campaign at a police summit in Dubai on Wednesday. [23]

    • Mohamed Al Kuwaiti
    • Power targeted
    • Water targeted
    • Hospitals targeted

    Why it matters

    Attackers also tried to trick airport passengers into paying money or changing flights. [23]

    Think it through

    1. 1

      Why erase data rather than ask for money?

      The aim was to undermine public trust in the UAE's leaders and systems, Al Kuwaiti said. [23]

    2. 2

      Why are the attacks hard to trace?

      AI agents made it hard to trace where the attacks came from, he said. [23]

    3. 3

      So who is the UAE relying on?

      He said international co-operation was essential to find those responsible. [23]

    How it works

    Security studies calls an attack aimed at damage, not money, sabotage.

    Wipers aimed at the UAE's power and water were sabotage: they erase data and ask for nothing. [23]

    Seen it before

    Someone who scribbles over your homework wants you to fail, not to sell it back.

    The truth

    • The 800,000 is a peak, from the height of the conflict, not an everyday rate. [23]
  8. 10

    US points to Iran over France attacks

    What happened

    • US intelligence agencies have reached a "low confidence" view that Iran was behind recent attacks in France. [24]
    • Attacks in October hacked hundreds of supermarket fridge parts and targeted cold stores for France's blood supply. [13]
    • Millions of dollars of food was spoiled. [13]
    • The officials tied it to Iran in general, not to a named group such as the Revolutionary Guards. [24]

    Why it matters

    Officials think Iran is also probably behind similar attacks elsewhere in Europe. [24]

    Think it through

    1. 1

      What does "low confidence" mean?

      Some intelligence points that way, but it is fragmented and could change with more information. [24]

    2. 2

      So who settles who did it?

      More evidence would; the other European cases are at an early stage, the officials said. [24]

    How it works

    Intelligence analysts call naming an attacker attribution, and they grade how sure they are.

    Here the US gave its attribution of the France attacks the grade "low confidence". [24]

    Seen it before

    A teacher who sees paint on one pupil's sleeve suspects them, but did not see them do it.

    The truth

    • Three US officials described it on condition of anonymity, so the assessment itself is not public. [24]
  9. 11

    Ransomware suspects arrested in Ukraine

    What happened

    • Ukraine's police, working with the FBI, arrested 4 suspected members of a ransomware gang. [25]
    • Investigators tie the group to $40.5 million in ransoms from at least 28 companies. [25]
    • The group had at least 27 members, from Ukraine, Kazakhstan, Russia and other countries. [25]
    • Quantum Health, a US company, paid $5 million in June after a $15 million demand, court records say. [25]

    Who is involved

    Ransomware locks or steals a company's files, and the gang demands money to give them back. [25]

    Why it matters

    Its victims included oil and gas, drug, factory and finance firms. [25]

    Think it through

    1. 1

      Where did this gang come from?

      Some members came from Conti, a big ransomware gang that shut down in May 2022. [25]

    2. 2

      Does arresting 4 stop it?

      Police arrested 4, but the wider network has more than 60 members in 8 places. [25]

    3. 3

      So who keeps paying?

      The victims do: 28 companies paid, and police say they are still counting the damage. [25]

    How it works

    Criminologists study gangs like this as networks: when one breaks up, its members regroup elsewhere.

    Conti's members split into smaller gangs, and some ended up in this network. [25]

    Seen it before

    When a big band splits, its players turn up in new bands playing the old songs.

    The truth

    • The victims and sums come from Ukrainian court records, reported by LeMagIT, not from the companies. [25]

    Good news

    • Police in Ukraine, working with the FBI, arrested 4 suspects tied to $40.5 million in ransoms after a two-year investigation.
  10. 12

    Scattered Spider hacker jailed 6 years

    What happened

    • Tyler Buchanan, 24, from Dundee in Scotland, was jailed for 6 years by a US court. [27]
    • The gang texted hundreds of fake login links to staff at phone and tech companies. [27]
    • They moved victims' phone numbers to their own devices to get past login codes. [27] Then they stole at least $11.6 million in crypto. [27]
    • Over the years the overlapping gangs hit big firms including Coinbase, DoorDash, Mailchimp and Nvidia. [28]
    • He is the last of 5 defendants in the US case to be sentenced. [27]

    Who is involved

    He was a leader of Scattered Spider, a loose group of young English-speaking hackers. [26]

    • Tyler Buchanan
    • Coinbase hit
    • DoorDash hit
    • Nvidia hit

    Why it matters

    Prosecutors said the thefts had life-altering effects for many victims. [26]

    Think it through

    1. 1

      How did texts beat company security?

      The texts led to fake pages that looked like the staff's own employers' websites. [26]

    2. 2

      How do the 5 sentences compare?

      Noah Urban got 10 years, Buchanan 6 and Joel Evans 2. [27] Two others got 45 months each. [27]

      prison terms in the US case
    3. 3

      So will victims get money back?

      A court hearing on 11 December will decide what Buchanan must repay. [27]

    How it works

    Security engineers call it SIM swapping: moving someone's phone number to a SIM card you control. [26]

    With the numbers moved, Scattered Spider received the login codes, so the SIM swap beat the second check. [27]

    Seen it before

    If the post office sends your mail to a thief, the thief gets your new bank card.

    The truth

    • The theft total grew from $8 million at his guilty plea to $11.6 million now. [27]

    Good news

    • All five defendants in the US Scattered Spider case have now been sentenced, the last to 6 years in prison.
  11. 13

    Coinbase stops a Musk deepfake scam

    What happened

    • Coinbase, a big crypto exchange, says it blocked $4.7 million in transfers to a deepfake scam. [29]
    • Scammers ran a fake YouTube stream of Tesla's "flying car" show, with an AI-made video of Elon Musk. [29]
    • The fake Musk promised to double any crypto that viewers sent. [29]
    • Hundreds of Coinbase customers tried to send money over about 2 hours. [29]
    • Its system stopped about 91% of the attempted transfers. [29]

    Why it matters

    The victims sent the money themselves, so checks built for stolen accounts would see nothing wrong. [29]

    Think it through

    1. 1

      How did Coinbase spot it?

      Its software watches outgoing transfers for signs like giveaway promises and brand-new wallets. [29]

    2. 2

      What happens when a transfer is flagged?

      An AI agent chats with the customer and explains the signs of the scam. [29]

    3. 3

      So who stands between the scam and the money?

      The exchange does, because crypto that is sent cannot be pulled back. [29]

    How it works

    Fraud investigators call this authorised push payment fraud: the victim sends the money themselves.

    Coinbase's customers were logged in and pressed confirm. [29] That is why authorised push payment fraud slips past checks for stolen accounts. [29]

    Seen it before

    A friend who grabs your arm before you hand lunch money to a stranger's too-good offer.

    The truth

    • Coinbase has not said how much reached the scammers, and the 91% is its own figure. [29]

    Good news

    • Coinbase stopped about 91% of the transfers its customers tried to send to a fake Elon Musk, keeping $4.7 million away from the scammers.
  12. 14

    AI debt collector exposed 6.6 million accounts

    What happened

    • Two flaws in CreditNirvana, an Indian AI debt-collection service, exposed 6.6 million accounts. [30]
    • They also exposed about 315,000 recordings of its AI call centre's calls. [30]
    • The data included credit reports, ID cards and legal papers. [30]
    • Eaton Zveare, an independent researcher, found the flaws; they were fixed in June and made public in October. [30]

    Why it matters

    People behind on debts had their private money troubles open while the flaws were live. [30]

    Think it through

    1. 1

      Who uses this service?

      Some of India's largest fintech firms use it to collect debts automatically, Zveare says. [30]

    2. 2

      Could attackers have run its AI callers?

      One flaw could have let an attacker direct the platform's AI agents, Wired reports. [30]

    3. 3

      So who pushed it to be fixed?

      Zveare worked with CERT-In, India's computer emergency team, which rated one flaw critical. [30]

    How it works

    Security engineers call it broken access control: a system fails to check who may see which records.

    CreditNirvana's broken access control could let an attacker open sessions as other users, Zveare found. [30]

    Seen it before

    A locker that opens with any pupil's code is no locker at all.

    The truth

    • CreditNirvana and its owner Perfios did not answer Wired, so whether others used the flaws is unknown. [30]
  13. 15

    Banks put AI agent risk on customers

    What happened

    • Citibank changed its terms: customers who let AI tools into their accounts carry the risk. [31]
    • That includes losses if the AI tool is hacked or leaks their login details. [31]
    • JPMorgan and Bank of America set similar rules last month, and PNC changed its terms this week. [31]
    • AI agents can be given bank logins to track spending and even pay bills. [31]

    Who is involved

    • Citi this week
    • JPMorgan last month
    • PNC this week

    Why it matters

    Senators say some banks cannot tell a rogue AI agent from one acting for a customer. [36]

    Think it through

    1. 1

      Why are banks doing this now?

      Banks see the tools as unreliable, TechCrunch reports. [31] They also see them as rivals for their customers. [31]

    2. 2

      Is any AI firm playing it safe?

      Anthropic's new personal finance tool is read-only and cannot move money. [31]

    3. 3

      So who holds the risk?

      The customer does under the new terms, even when it is the AI firm that gets hacked. [31]

    How it works

    Lawyers call this allocating liability: a contract says in advance who pays when something goes wrong.

    Citi's new terms allocate liability for an AI agent's actions to the customer. [31]

    Seen it before

    Lend your bike to a friend under a family rule that says any damage is yours.

    The truth

    • Citi says it wants customers to understand the risks. [31] Banks also see AI tools as a threat to their fees, TechCrunch reports. [31]
  14. 16

    AI bug hunters barely agree

    What happened

    • Three AI models checked the same open-source code and agreed on just 1 of 121 possible flaws. [32]
    • Echo, a software security firm, ran the test on three open-source code libraries. [32]
    • Of the 121 issues, 100 came from one model, 20 from two and 1 from all three. [32]

    Who is involved

    The models were Anthropic's Claude Mythos 5.1, OpenAI's GPT-5.6 Sol and Z.ai's GLM-5.3. [32]

    • Claude Mythos 5.1 Anthropic
    • GPT-5.6 Sol OpenAI
    • GLM-5.3 Z.ai

    Why it matters

    Echo's researcher Yuval Saggie called the race to find the best model "the wrong question". [32]

    Think it through

    1. 1

      Do they agree on how bad a flaw is?

      No: one flaw was rated Critical, High, and not a bug at all by the three models. [32]

    2. 2

      How many findings held up?

      Of 55 findings tested with a working proof, 42 passed. [32]

      42 of 55findings that passed testing

    3. 3

      So what decides which flaws get found?

      Largely which model a team uses: most issues had only one finder in Echo's test. [32]

    How it works

    Software testers call this coverage: each method checks some of the code's paths and misses others.

    Each model's coverage was different, so 100 of the 121 issues had a single finder. [32]

    Seen it before

    Three friends searching one messy room each find different lost socks.

    The truth

    • It is one firm's test of three libraries, and Echo says the costs reflect its own access terms. [32]
  15. 17

    Old Apache blog server under attack

    What happened

    • Attackers are using a critical flaw in Apache Roller, an open-source blogging server, to run code on servers. [33]
    • The flaw lets a stranger run code without logging in. [33]
    • Apache says the same flaw, first listed only for version 6.1.5, reaches back to version 4.0.0. [33]

    Who is involved

    Roller is built on Java and has a small market share. [33]

    Why it matters

    Owners of older Roller versions may have thought the first notice did not apply to them. [33]

    Think it through

    1. 1

      Can owners just switch the feature off?

      No: turning it off in the admin screen is not enough, Apache warns. [33]

    2. 2

      So who has to act?

      Each Roller owner, by upgrading to 6.1.6 or editing a settings file by hand. [33]

    How it works

    Security engineers call this deserialization of untrusted data: a server rebuilds objects from a stranger's message.

    Roller did that deserialization before checking who sent the message, so the check came too late. [33]

    Seen it before

    Opening a parcel before checking who sent it.

    The truth

    • Apache has not said who is attacking or how many servers are hit. [33] That decides whether this is a few blogs or a wide campaign. [33]
  16. 18

    Puppet flaw attacked; CISA lists 4 more

    What happened

    • Attackers are exploiting a critical flaw in Puppet, a tool companies use to set up and manage their computers. [34]
    • It scores 9.4 out of 10 for danger. [34]
    • Perforce, Puppet's maker, patched it on 3 November, and attacks were flagged on 4 November. [34]
    • CISA, the US cyber defence agency, added it to its list of flaws under attack on 17 November. [34]
    • On 18 November CISA added 4 more, including flaws in BerriAI's LiteLLM and Vanderbilt University's REDCap. [35]

    How it unfolded

    1. 3 November fix released
    2. 4 November attacks flagged

    Why it matters

    A taken-over Puppet server could change the software it delivers to every machine it manages, Perforce warns. [34]

    Think it through

    1. 1

      How do attackers get in?

      They need a certificate from Puppet's own signing system, which some setups hand out automatically. [34]

    2. 2

      So who must move fastest?

      US federal agencies, which CISA tells to fix listed flaws within two weeks. [34]

    How it works

    Security engineers call the part that issues trusted certificates a certificate authority. [34]

    Puppet's certificate authority decides which machines are trusted. [34] Steal its key, and an attacker can issue trusted certificates. [34]

    Seen it before

    Whoever holds the school's stamp can make any pass look real.

    The truth

    • The fix came a day before attacks were flagged, so the danger sits with servers not yet updated. [34]
  17. 19

    Oracle and Chrome ship hundreds of fixes

    What happened

    • Oracle's November security release covers 948 flaws, SecurityWeek counted. [37]
    • 62 of Oracle's fixes are for critical flaws. [37]
    • They include the fix for Java4Shell, the serious Java flaw found earlier this week. [37]
    • Google's Chrome 158 fixed 204 flaws, 6 of them critical. [38]

    Why it matters

    Every company running Oracle software or Chrome has another round of updates to test and install. [37]

    Think it through

    1. 1

      Are any of these being attacked?

      Google mentions no Chrome flaw being used in attacks. [38]

    2. 2

      Who is finding Chrome's bugs?

      Google's own teams found 145 of them; outside researchers were paid $11,000 for 5. [38]

    3. 3

      So who carries the load?

      The people and IT teams who must install them; users are advised to update as soon as possible. [38]

    How it works

    IT teams call it patch management: choosing which fixes go in first, then testing and installing them.

    Just over 300 of Oracle's fixes close flaws usable remotely without a login, the kind patch management ranks high. [37]

    Seen it before

    A teacher with a pile of marking reads the essays due today first.

    The truth

    • SecurityWeek counted 948 unique flaws in Oracle's notes, but only 640 appear in Oracle's risk tables. [37]
  18. 20

    Another critical PHP update due

    What happened

    • The PHP Foundation will release a "critical" security update on Thursday between 14:00 and 14:30 UTC. [39]
    • The worst flaws need unusual settings to be used, the foundation says. [39]

    Who is involved

    PHP is a programming language running on about 70% of websites. [39]

    How it unfolded

    Last month's PHP2Shell flaws went on CISA's list of attacked flaws on 26 October. [39]

    1. Wednesday advance notice
    2. Thursday 14:00 UTC the update
    3. Last month PHP2Shell fixed
    4. 26 October on CISA's attacked list

    Why it matters

    Website owners face another round of updates in a week already crowded with fixes. [39]

    Think it through

    1. 1

      Why announce it before releasing it?

      So users can "prepare to upgrade promptly", the foundation said. [39]

    2. 2

      What happened after the last fixes?

      Suspected Chinese attackers used PHP2Shell to take over 25 servers, including Palo Alto firewalls, Dream researchers say. [39]

      25servers taken over with PHP2Shell

    3. 3

      So who sets the clock?

      The foundation sets the release time. [39] Attackers then work backwards from the fix, though PHP2Shell's took them longer than expected. [39]

    How it works

    Security researchers call it patch diffing: comparing old and new code to see what a fix closed.

    After PHP2Shell, attackers took longer than some researchers expected to reverse engineer the fixes. [39]

    Seen it before

    Seeing which floorboard a builder replaced tells you where the hole was.

    The truth

    • One flaw in Thursday's release was judged too unlikely in October to need an emergency fix. [39]
    • Nobody knows yet

      How urgent Thursday's PHP update really is.

      The foundation says the critical issues need unusual setups and has given few details before the release. [39]

  19. 21

    Chinese spy tools rebuilt after exposure

    What happened

    • A suspected Chinese spy group rebuilt its EggStreme hacking tools after Bitdefender, a security firm, exposed them. [40]
    • A changed part appeared in a build made 12 days after the report. [40]
    • The operators swapped their codes so detections built for the old ones stopped working. [40]

    How it unfolded

    Bitdefender first reported them in September 2025, after an attack on a Philippine military firm. [40]

    Why it matters

    Philippine government bodies were the main targets, and possibly Taiwan too. [40]

    Think it through

    1. 1

      Did being exposed stop them?

      No: they changed tools within weeks but kept 4 core methods, Bitdefender says. [40]

    2. 2

      So what gives them away now?

      Those methods, such as getting real programs to load bad code, stayed the same. [40]

    How it works

    Threat analysts separate tools, which are cheap to change, from methods, which are costly to change.

    EggStreme's operators "spent at the tools level and kept their methods", so defenders can still watch the methods. [40]

    Seen it before

    A burglar can buy new gloves, but still climbs in through back windows.

    The truth

    • Bitdefender is highly confident only about the Philippines; Taiwan is a lead, not an established fact. [40]
  20. 22

    Senators ask for an AI defence plan

    What happened

    • US senators Maria Cantwell and Mike Rounds asked the US government for a national effort against AI-enabled cyber threats. [36]
    • They wrote that AI helps attackers find and use software weaknesses faster than many organisations can respond. [36]
    • Their letter went to Jay Clayton, who chairs the US government's task force on AI. [36]
    • They want faster defensive AI tools and security standards for AI agents. [36]

    Who is involved

    • Maria Cantwell Democrat, Washington
    • Mike Rounds Republican, South Dakota

    Why it matters

    Testing and installing fixes across connected systems can take weeks or months, they wrote. [36]

    Think it through

    1. 1

      What prompted the letter?

      It cites recent attacks in Japan where AI found obscure bugs and powered break-ins. [36]

    2. 2

      Who do they want in the room?

      AI developers, security firms, critical infrastructure operators, NIST and CISA, among others. [36]

    3. 3

      So who has to act?

      Clayton's task force, which the senators say must speed up building defence tools. [36]

    How it works

    Policy experts call this a coordination problem: no single firm can fix a risk that every system shares.

    The senators' answer to the coordination problem is one meeting of developers, firms and agencies. [36]

    Seen it before

    A school fire drill only works if every class leaves at once.

    The truth

    • It is a letter, not a law, so the task force decides what to do with it. [36]
  21. 23

    UK audit office faults risk handling

    What happened

    • The National Audit Office, which checks UK government spending, says government falls short on managing risk. [41]
    • 58% of the risk staff it polled said they are brought in too late to shape decisions. [41]
    • The number of chief risk officers in government has more than doubled in 10 years. [41]
    • Leaders often do not understand the cyber risks listed in their own risk registers, it found. [41]

    Why it matters

    Fast risks like the recent PHP flaws can disrupt public services, the report says. [41]

    Think it through

    1. 1

      Do risk teams change any decisions?

      The NAO found limited evidence that their input changed major decisions. [41]

    2. 2

      Are the risks within agreed limits?

      In June, most departments said most of their main risks were outside their stated tolerance. [41]

    3. 3

      So who holds the power?

      Senior leaders do; low seniority can keep risk heads out of the rooms where decisions are made. [41]

    How it works

    Risk managers call the amount of risk an organisation will accept its risk appetite. [41]

    Only 47% of those polled said their department's risk appetite actually shapes decisions. [41]

    Seen it before

    A family agrees a spending limit but never checks the bill.

    The truth

    • Most findings come from a poll of risk professionals, not from tracking actual decisions. [41]
  22. 24

    Judge lets Google buy Spirit staff data

    What happened

    • A US bankruptcy judge approved the sale of Spirit Airlines' former staff data to Google. [42]
    • The data includes about 100 million emails and half a billion Teams messages. [42]
    • Google plans to use it to train AI. [42]
    • Google agreed to leave out health, drug-test and discipline records, and not to identify anyone. [42]

    How it unfolded

    Spirit, a US airline, shut down in May and agreed to sell the data for $10 million. [42]

    Why it matters

    Former Spirit workers' messages become material for AI agents that do office work, experts say. [42]

    Think it through

    1. 1

      Why did the union drop its objection?

      Google promised to strip sensitive records and let an independent assessor check the cleaned data. [42]

    2. 2

      Could people still be traced?

      A privacy start-up's chief warns that each time the data is copied or breached, "someone always finds an opening". [42]

    3. 3

      So who checks the promise?

      An independent assessor, chosen with the flight attendants' union, reviews the cleaned data. [42]

    How it works

    Data protection experts call removing what points to a person de-identification. [42]

    Here de-identification has to cover 100 million emails, which the judge said meets data protection law. [42]

    Seen it before

    Blurring faces in a class photo still leaves the uniforms that show the school.

    The truth

    • The promise not to reidentify anyone rests on Google's word and the assessor's review. [42]
  23. 25

    Palo Alto buys AI firm for $4.5bn

    What happened

    • Palo Alto Networks, a big security company, will buy Applied Compute, an AI firm, for $4.5 billion in shares. [43]
    • Palo Alto plans to offer it to its security customers. [43]
    • The deal should close in Palo Alto's 2027 fiscal third quarter, if regulators agree. [43]

    Who is involved

    Applied Compute helps companies adapt open AI models to their own data and run them. [43]

    How it unfolded

    1. Now deal agreed
    2. 2027 expected to close

    Why it matters

    Its customers could run AI models in public clouds or in their own countries' systems. [43]

    Think it through

    1. 1

      Did Palo Alto test it first?

      It tried Applied Compute as a customer, and its Unit 42 team checked the platform's security. [43]

    2. 2

      So who gains?

      Applied Compute's chief, Yash Patil, becomes a co-chief AI officer at Palo Alto. [43]

    How it works

    Business strategists call this moving into an adjacent market: selling the next thing your customers need.

    Palo Alto moves into that adjacent market, a role in how firms build AI, not only guard it. [43]

    Seen it before

    A bike-lock shop that starts selling bikes.

    The truth

    • The price is in shares and "subject to adjustments", so the final value can change. [43]
  24. 26

    Windows Server 2028 ships locked down

    What happened

    • Microsoft announced Windows Server 2028, which comes with hardened security settings by default. [44]
    • Weak legacy settings, such as unauthenticated network calls, are switched off from the start. [44]
    • Its encryption uses a key exchange meant to resist future quantum computers. [44]
    • Admins can relax specific settings if they need to. [44]

    Why it matters

    Microsoft says the defaults cut the work of making a server secure "from day 1". [44]

    Think it through

    1. 1

      What exactly is switched off?

      Guest sign-ins and unauthenticated network calls are off, and Kerberos logins are preferred over older NTLM. [44]

    2. 2

      So who decides how safe each server is?

      Microsoft sets the starting point; each admin decides what to loosen. [44]

    How it works

    Security engineers call this secure by default: safe settings stay on until someone chooses otherwise.

    Windows Server 2028 is secure by default, so an admin must decide to relax a setting. [44]

    Seen it before

    A new phone that arrives with its screen lock already on.

    The truth

    • The defaults hold only while admins keep them, since the new baseline lets them relax settings. [44]
  25. 27

    Turkish pirate TV site shut down

    What happened

    • Dizigom, one of Turkey's biggest pirate TV and film sites, has shut down. [45]
    • It had about 21 million visitors a month. [45]
    • Its operators agreed to stop and hand 16 web addresses to ACE, a studio-backed anti-piracy group. [45]

    How it unfolded

    Turkish police and INTERPOL helped in an investigation that began in 2023. [45]

    1. 2023 investigation begins
    2. Now site shut

    Why it matters

    Its addresses now send visitors to a page listing legal streaming services. [45]

    Think it through

    1. 1

      Why a deal, not a raid?

      The settlement secured a permanent shutdown and the web addresses, ACE says. [45]

    2. 2

      So who holds the power here?

      ACE, backed by more than 50 studios and TV networks, including Disney and Netflix. [45]

    How it works

    Lawyers call this a settlement: both sides agree terms instead of fighting to a verdict. [45]

    Dizigom's operators settled once criminal cases in Turkey had advanced. [45]

    Seen it before

    Two neighbours agree to move a fence instead of going to court.

    The truth

    • Pirate sites often reappear, and ACE says it targets the people behind them to stop that. [45]
  26. 28

    Hospital staff fired for reading a patient's file

    What happened

    • Three staff at Addenbrooke's Hospital in Cambridge were dismissed for looking at a patient's records. [46]
    • Staff with no part in her care read her notes and saw her photos. [46]
    • She died of breast cancer in June 2025, aged 28, and her sister is now speaking out. [46]

    Who is involved

    Gemma Sabatino was in critical care there during a mental health crisis in late 2023. [46]

    Why it matters

    NHS trusts in England logged more than 3,000 cases of staff wrongly opening patient data from 2021 to 2026. [46]

    Think it through

    1. 1

      How was it found out?

      The family pushed for an inquiry, and the system showed every click, download and print. [46]

    2. 2

      So who guards the records now?

      The trust says it has tightened its checks, and it reported itself to the ICO in 2024. [46]

    How it works

    Information security calls a record of who opened what an audit trail.

    The hospital's audit trail showed who opened Gemma's file once her family complained. [46]

    Seen it before

    A library log shows who borrowed a book, but only if someone looks.

    The truth

    • The 3,000 counts recorded cases only; snooping nobody spotted is not in it. [46]

Across the beats