Attackers use ScreenConnect, a remote-support tool, to reach 286 computers at 47 organisations through six IT firms

The stories

25 on this beat today.

Every story told the same way: what happened, who is involved, why it matters, how it works, the truth and the good news.

1. 01

   Attackers use ScreenConnect, a remote-support tool, to reach 286 computers at 47 organisations through six IT firms

   286

   computers with a rogue copy of ScreenConnect, and counting

   spread across 47 organisations, Huntress says [1]

   6

   IT service firms broken into so far

   each runs computers for many client companies [1]

   400

   IT firms' setups where ConnectWise switched off 5 add-ons

   a precaution while it investigates [1]

   What happened

   • Attackers used ScreenConnect, a remote-support tool, to put rogue copies of it on 286 computers at 47 organisations. [1]
   • They got in through 6 managed service providers, firms paid to run other companies' computers. [1]
   • Huntress, a security firm, says the attacks began around 17:47 UTC on Wednesday, and the count is still rising. [1]
   • The attackers used ScreenConnect's own Run Command feature to install their rogue copies on client computers. [1]
   • Each rogue copy gave them a way in that the IT firm did not control. [1]
   • Spike Brandt of Huntress called it "abuse of legitimate functionality", not a flaw in the software. [1]
   • ConnectWise switched off 5 add-ons from other companies across about 400 IT firms' setups, as a precaution. [1]
   • It suspects the attackers installed more remote-access software through those add-ons, which partners pick from its built-in store. [1]
   • Customers who run ScreenConnect on their own servers must switch the add-ons off themselves. [1]
   • Huntress says its data points to stolen logins or hijacked sessions at the IT firms. [1]
   • ConnectWise has reports that technicians' own computers were infected while connected to hacked machines. [1] It is still checking them. [1]
   • Numbered tags the attackers used suggest other IT firms were broken into earlier, Huntress says. [1]

   Who is involved

   ScreenConnect is made by ConnectWise, a US company, and lets an IT worker control a customer's computer from far away. [1][2]

   Service firms use it to fix and manage many client computers from one place. [1]

   • ConnectWise the US company that makes ScreenConnect; it confirmed the break-ins and switched off 5 add-ons
   • ScreenConnect remote control of a computer
   • Huntress a security company that watches clients' computers; it found the attacks on Wednesday
   • Managed service providers firms paid to run other companies' computers; 6 of them were broken into

   How it unfolded

   Last month ConnectWise found attackers in its Cloud Backup service on 14 October, using a stolen login. [3][2]

   Microsoft blamed Storm-0940, a China-linked group, which tried to sign in to more than 2,250 companies' Microsoft 365 accounts. [3]

   1. 14 Oct ConnectWise finds attackers in its Cloud Backup service [3]
   2. Now ScreenConnect attacks begin
   3. 21 Oct Microsoft blames Storm-0940, a China-linked group [3]
   4. 11 Nov, 17:47 UTC rogue ScreenConnect copies start appearing [1]
   5. 11 Nov, night ConnectWise confirms the break-ins and switches off 5 add-ons [1]

   Why it matters

   Each IT firm looks after many client companies, so 6 break-ins reached 47 organisations. [1]

   Insurer At-Bay's 2021-2024 claims showed remote access tools were attackers' most common target for getting ransomware in. [4]

   How it works

   Security analysts call this living off the land: an attacker uses tools already trusted on a network.

   The rogue copies were ScreenConnect itself, a program IT teams install on purpose. [1] Huntress says attacks starting with real remote-management tools rose sharply in 2026. [1]

   Seen it before

   A burglar in a delivery uniform walks past a guard who only stops strangers.

   Why rules for hacking tools check the buyer, not the tool

   Arms-control experts call tools like these dual-use: they break in for police and criminals alike, so the rules vet the buyer instead.

   The twist

   A hacking tool works the same whoever holds it. So the only place a rule can act is the sale: who may buy it, and for what.

   The picture

   Countries at the meeting43 countries

   Countries that adopted the rules14 countries

   Step by step

   1. 1 A tool is built to break in, for a lawful job
   2. 2 It works the same for anyone who holds it
   3. 3 So a rule cannot ban what the tool does
   4. 4 Instead the rule checks who buys it, and why
   5. 5 That check needs a seller or a licence to stand on

   The same force, elsewhere today

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

   • AI testing tool turned on cloud accounts

     AntiFox is a free download with no seller to vet anyone, so its makers' ban on attacks has nothing in the code to enforce it.

   • The ScreenConnect attacks

     The remote-support tool IT firms use to fix clients' computers did the same job for the attackers on 286 of them.

   • Attackers use a Cisco phone-system flaw

     Attackers removed the limits from an AI model and used it to attack a phone company, so the rules its makers built in did not hold.

   • Fortinet fixes nearly 100 flaws at once

     Fortinet used frontier AI to find nearly 100 flaws in its own products, and says AI also speeds up attacks once a flaw is public.

   Where you've seen this

   Cold medicine at a chemist

   some kinds are sold only from behind the counter, because they can be made into illegal drugs

   Farm fertiliser

   in some countries sellers must record buyers, because the same bags can make a bomb

   Locksmith tools

   the picks that open a locked-out family's door also open a burglar's target, so some places license locksmiths

   The catch

   A check at the sale works only where there is a seller. A free tool anyone can download, like AntiFox, has no counter to stand behind.

   And the whole of it

   The remote tools that let an IT firm fix a client's computers from another city let a stranger in too. Many office computers are looked after this way, by firms and tools the people using them never see.

   The truth

   • What is really going on Attackers reached 286 computers through ScreenConnect, a tool six IT firms trusted, and ran AntiFox, a free testing tool, against Amazon cloud accounts. [1][6]
   • ConnectWise only suspects the add-ons, and says having them does not prove a network was hit. [1]
   • Who pays The 47 organisations reached through ScreenConnect — Switching off the add-ons removes nothing already installed, so each computer has to be checked. [1]
   • Nobody knows yet

     How the ScreenConnect attackers first got into the six IT firms.

     Huntress says it has no information on the first machine, and its data only points to stolen logins or hijacked sessions. [1]

   • Nobody knows yet

     How many more IT firms were hit before Wednesday.

     Numbered tags in the attack suggest earlier break-ins that nobody has counted yet. [1]

   • Nobody knows yet

     Whether technicians' own computers were infected during sessions.

     ConnectWise has reports of it but says they are still being confirmed. [1]

   What comes next

   ConnectWise has told partners not to turn the add-ons back on until it issues new guidance. [1]

1. 02

   Gang claims Maersk data, wants $5 million

   What happened

   • Maersk, one of the world's biggest container shipping firms, confirmed that hackers stole data from an old system. [14]
   • A gang called WhiteArmor claims it took about 100GB of data and wants $5 million in bitcoin. [14]
   • It gave Maersk 30 days to pay, and says the files hold nearly 30,000 driver records and plaintext passwords. [14]
   • Maersk says the break-in is limited to one local system, and it expects no disruption to operations. [9]

   How it unfolded

   Maersk says the system came with a firm it bought. [14] The gang names LF Logistics, which Maersk bought from Li & Fung in 2022. [14]

   In 2017 the NotPetya attack forced Maersk to rebuild about 4,000 servers and 45,000 PCs. [9]

   1. 2022 Maersk buys LF Logistics
   2. Now its old system is breached

   Why it matters

   Maersk says its customers can keep booking and managing shipments as usual. [14]

   How it works

   Security researchers call this extortion by leak: steal files, then threaten to publish them unless paid.

   WhiteArmor listed Maersk on its leak website and threatened to release the 100GB. [14]

   Seen it before

   A thief who copies your diary, then asks for money not to post it.

   The truth

   • Why it works on people A company's first statement that a breach is limited to one system sounds like an ending, though the inquiry has only begun. [9]
   • Only the gang has described the stolen data, and Maersk has not said which system was hit or when. [14]
   • Who gains Ransomware gangs like WhiteArmor — Naming Maersk in public with a 30-day deadline puts pressure on before Maersk has finished its inquiry. [14]
   • Who pays Drivers named in the files WhiteArmor claims — Maersk has warned people to check any unexpected message about the incident through their usual Maersk contact. [9]
   • Nobody knows yet

     What WhiteArmor really took from Maersk.

     Maersk has not named the data types or said when the intrusion happened. [14]

2. 03

   Poland switches off its online state services

   What happened

   • Poland switched off ePUAP, the website people use to send official papers to the state, after a cyberattack. [15]
   • Digital Affairs Minister Krzysztof Gawkowski said the attack was "detected and stopped". [15]
   • A ministry spokesman said they turned the site off as a precaution, not because the attack broke it. [15]
   • Government email boxes for citizens are down, and the Trusted Profile, a digital ID for signing documents, may be affected. [15]

   How it unfolded

   In September, a breach at MyDr, a medical software firm, may have reached data on about 19 million people. [15]

   Why it matters

   Citizens and businesses that file official papers online cannot do so until ePUAP returns. [15]

   How it works

   Incident responders call this containment: cut a system off first, then look for damage.

   Poland switched ePUAP off as a precaution, and says it returns once security experts finish their checks. [15]

   Seen it before

   A shop shuts its doors while it checks whether the alarm was a real break-in.

   The truth

   • Who pays People in Poland who use the Trusted Profile digital ID — They may be unable to sign official documents online while the ministry's checks go on. [15]
   • Nobody knows yet

     Who attacked Poland's ePUAP, and how.

     Poland's digital affairs ministry has disclosed neither, and has given no date for the site's return. [15]

3. 04

   N-able fixes the worst of 33 posted flaws

   What happened

   • N-able released a fix on Tuesday for the worst of 33 flaws a researcher posted with attack code. [16]
   • The fix covers 7 flaw numbers and switches off N-central's remote-control feature for now. [16]
   • All hosted N-central servers have been updated, N-able says. [5]
   • Security firm Defused saw attackers probing N-central servers for the posted flaws within 24 hours. [5]

   Who is involved

   N-able makes N-central, which IT firms use to watch and manage clients' computers from one screen. [5]

   • N-able makes N-central
   • N-central manages many computers

   Why it matters

   N-able says its tools are used at more than 500,000 organisations in over 140 countries. [5]

   The truth

   • Researcher Sean Bales told Infosecurity he warned N-able about none of these flaws before posting them. [16]
   • Nobody knows yet

     How the 7 fixed N-able flaw numbers match the 33 posted flaws.

     Neither the release notes nor N-able's statement say, and attack code for all 33 is still public. [16]

   Good news

   • Shadowserver now counts just over 148 N-central servers open to the internet, down from 1,338 on Monday. [5]

   What comes next

   N-able says a follow-up fix will come. [16] WatchTowr, a security firm, says it will not be the last. [17]

4. 05

   Attackers use a Cisco phone-system flaw

   What happened

   • CISA, the US cyber-defence agency, says attackers are using a flaw in Cisco BroadWorks, phone companies' calling software. [7]
   • The flaw lets anyone read a server's settings files by sending a crafted message, with no login. [7]
   • Cisco fixed it in August, and CISA listed it as under attack on 10 November. [18]
   • Malbear Labs, a security firm, says a Chinese-speaking group used the flaw with a hacked-open AI model. [7]
   • The group went after a European phone company and at least 13 organisations using its services. [18]
   • CISA also listed a flaw in Openwave's webmail software, which phone companies offer to customers, as under attack. [18]

   How it unfolded

   1. Aug Cisco fixes the flaw
   2. 10 Nov CISA lists it as attacked

   Why it matters

   The attackers went after customers' billing and device records, Malbear Labs says. [18]

   How it works

   BleepingComputer describes it as an XML external entity flaw, where a data reader also fetches files a message points to. [7]

   Cisco says BroadWorks allowed those outside references by default, so a crafted message made it open its own files. [7]

   Seen it before

   A clerk who fetches any file named on a form, even one the visitor should never see.

   Attackers removed the limits from an AI model and used it to attack a phone company, so the rules its makers built in did not hold.

   The idea, in story 01: Why rules for hacking tools check the buyer, not the tool →

   What comes next

   US federal agencies must patch both flaws by 24 November. [18]

5. 06

   Iran denies French supermarket fridge attacks

   What happened

   • Iran's foreign minister called France's claim that Iran hacked French supermarket fridges "baseless". [19]
   • France says they caused fires and tens of millions of euros of spoiled food. [19]
   • France plans new sanctions, and Paris prosecutors have opened a judicial investigation. [19]
   • A second wave in November hit over 100 stores in countries including Belgium, Sweden and the Netherlands. [19]

   Who is involved

   • France accuses
   • Iran denies

   How it unfolded

   France this week blamed Shahid Kaveh, a cyber unit of Iran's Revolutionary Guards. [19]

   The October attacks reset fridge controllers at more than 700 French food stores to high temperatures. [19]

   Why it matters

   Carrefour, Monoprix, E.Leclerc, Cooperative U and Intermarche were among the French chains hit. [19]

   The truth

   • Nobody knows yet

     What evidence ties the French fridge attacks to Iran's Shahid Kaveh unit.

     France says its investigation found the link, and Iran calls the claim baseless. [19]

6. 07

   Baker Tilly clients' data stolen at a supplier

   What happened

   • Baker Tilly, a big US accounting firm, is telling clients that a hacker took their data from a supplier's platform. [10]
   • Texas's attorney general lists 10,456 people affected in that state alone, with Social Security numbers exposed. [10]
   • Baker Tilly learned of it on 6 September and found whose data was involved on 28 October. [10]

   Who is involved

   The firm works in 147 countries and earns $6.8 billion a year. [10]

   How it unfolded

   1. 6 Sep breach found
   2. 28 Oct affected clients identified

   Why it matters

   Clients' dependents are affected too, and Baker Tilly found no reports of identity theft so far. [10]

   How it works

   Risk managers call this third-party risk: data is only as safe as each firm it is handed to.

   Baker Tilly found no sign of an intruder on its own network. [10] The stolen data sat on the supplier's system. [10]

   Seen it before

   Lending your bike to a friend means it is only as safe as their lock.

   The truth

   • What is really going on Maersk, Baker Tilly and Landmark Group have not said how their attackers first got in. [14][10][11]
   • Who pays Baker Tilly clients and their dependents — With Social Security numbers exposed, they are offered 24 months of credit monitoring through Experian. [10]
   • Nobody knows yet

     How many Baker Tilly clients are affected, and which supplier was breached.

     The notice names neither, and only one US state has published a count. [10]

7. 08

   Landmark's Shukran loyalty data reached

   What happened

   • Landmark Group, a big Dubai retailer, says someone got into the system behind Shukran, its loyalty points scheme. [11]
   • Names, contact details and dates of birth may be exposed. [11]
   • Card details were not stored on that system, and Shukran points are untouched, the company says. [11]

   Who is involved

   Landmark owns more than 20 brands, including Centrepoint, Max and Home Centre, across the Gulf and beyond. [11]

   How it unfolded

   IBM puts the average cost of a breach in the Middle East at $8 million this year. [11]

   Why it matters

   Customers who get a text from Landmark about the incident may be among those affected. [11]

   The truth

   • Nobody knows yet

     How many Shukran members' details were reached.

     Landmark has not given a figure or said how the intruder got in. [11]

8. 09

   First rules for sellers of hacking tools

   What happened

   • Governments from 14 countries adopted the first guidelines for companies that sell tools to break into phones and computers. [20]
   • Sellers are asked to vet buyers, build safeguards against misuse, and stop sales when serious misuse is a credible risk. [20]
   • The rules are voluntary, with no penalties, and 43 countries attended the meeting. [20]

   Who is involved

   The talks, called the Pall Mall Process, are led by Britain and France. [20]

   Signers include the US, Germany, Italy and Japan. [20]

   • Britain
   • France
   • US
   • Germany
   • Italy
   • Japan

   Why it matters

   France's foreign ministry says anyone can now buy "off the shelf" the ability to break into a phone. [20]

   It says such tools have been used to spy on journalists, officials, activists and businesses. [20]

   The truth

   • Who gains Britain and France — Some participants said last year that the two want to limit other states' access to hacking tools while backing their own spyware firms. [20]

   Good news

   • Governments from 14 countries, including Britain, France, the US and Japan, adopted the first shared rules asking sellers of hacking tools to vet their buyers.
9. 10

   AI testing tool turned on cloud accounts

   What happened

   • Attackers are running AntiFox, a free AI tool for testing defences, against websites and Amazon cloud accounts, Arctic Wolf says. [6]
   • In one break-in, the attackers used real Amazon logins to make hundreds of requests in about 30 minutes. [6]
   • Most attempts were blocked, but some got production secrets, the passwords a live system runs on. [6]

   Who is involved

   Arctic Wolf, a security firm, found the attackers' AntiFox copies running on Alibaba Cloud servers. [6]

   How it unfolded

   ARTEX, the tool AntiFox was copied from, went closed after it was used against South Korean banks. [6]

   Why it matters

   Arctic Wolf tells defenders to cancel the affected Amazon keys and change passwords stored in app files. [6]

   How it works

   AntiFox is a free download with no seller to vet anyone, so its makers' ban on attacks has nothing in the code to enforce it.

   The idea, in story 01: Why rules for hacking tools check the buyer, not the tool →

   The truth

   • AntiFox's makers ban attacks on live systems, but describe no technical control that enforces it. [6]
   • Who gains Attackers using free AI testing tools — AntiFox is on GitHub for anyone, and nothing in it stops an attack on a live system. [6]
10. 11

    Clues point to the man behind Qilin

    What happened

    • Brian Krebs, a security reporter, says clues point to Qilin's leader being a 36-year-old man in St Petersburg. [21]
    • Qilin is the busiest ransomware gang, claiming at least 365 victims in the past 3 months. [21]
    • Its hackers steal files before locking them, then threaten to publish them unless paid. [21]
    • Qilin pays the hackers who work for it at least 80% of each ransom. [21]
    • Flight records suggest the man has not flown outside Russia. [21]

    Why it matters

    Qilin's victims face a choice: pay, or see their stolen files posted online. [21]

    How it works

    Investigators call this open-source intelligence: building a picture from records others can search.

    Messages from a seized crime forum linked the leader's new name to older ones. [21] The trail led to an email address in leaked records of Russia's state services website. [21]

    Seen it before

    A nickname at school stops hiding you once someone links it to your real name.

    The truth

    • Krebs could not reach the man, and the trail runs through leaked databases and private intelligence firms. [21]
11. 12

    eMudhra cancels 8 more certificates

    What happened

    • eMudhra, an Indian firm that vouches for websites' identities, cancelled 8 more of its certificates. [22]
    • The total is now 16: 8 with keys so weak anyone could work them out, and 8 handled badly. [22]
    • A faulty script used in April 2025 to move keys between secure key boxes caused both problems. [22]
    • eMudhra missed an early warning in June 2025, when it fixed mismatched keys without asking why. [22]
    • Cisco and SonicWall have removed root certificates tied to eMudhra. [23]

    How it unfolded

    1. June 2025 mismatch fixed, not explained
    2. Oct a researcher finds weak keys
    3. 16 Nov eMudhra stops issuing
    4. 4 Dec KPMG review due

    Why it matters

    Browsers trusted these certificates, so a key anyone could work out could have let someone fake a real website. [22]

    The truth

    • eMudhra says it has found no evidence the weak keys were misused. [22]

    What comes next

    eMudhra stops issuing under its last certificate family by 16 November, and KPMG reports by 4 December. [22]

12. 13

    Leak shows China profiling chip talent abroad

    What happened

    • Leaked papers describe a Chinese state programme that scrapes social media to profile ethnic Chinese people abroad, Dark Reading reports. [24]
    • It runs from a 500-square-metre data lab at Jinan University in Guangzhou. [24]
    • One leaked report lists chip-industry figures, including Taiwanese executives and students, as people worth getting close to. [24]
    • It suggests approaching them through hometown groups, alumni groups and professional societies. [24]

    Who is involved

    NetAskari, a group that watches China, obtained the papers and shared its report with Dark Reading and Sky News. [24]

    The programme is run by the United Front Work Department, a Communist Party office for influence at home and abroad. [24]

    • NetAskari watches China
    • Dark Reading and Sky News got the papers

    Why it matters

    People can be profiled from what they post in public, without knowing a foreign state is watching. [24]

    The truth

    • Nobody knows yet

      How far newer AI has grown China's profiling programme.

      The leaked papers mention machine learning only in references from 2021. [24]

13. 14

    Fortinet fixes nearly 100 flaws at once

    What happened

    • Fortinet, which makes firewalls, released fixes for nearly 100 flaws on Tuesday, about 17 of them critical. [8]
    • The batch came 3 weeks after Fortinet said it had added frontier AI models to its own product testing. [8]
    • FortiMail, its email security product, got the most critical fixes. [8]
    • A researcher says one FortiClient flaw runs an attacker's code when a user visits a bad website. [8] It works only if FortiPAM, a Fortinet access tool, is on. [8]
    • Fortinet reports no attacks using any of these flaws. [8]

    Why it matters

    Fortinet says fixing before the details go public gives customers time to update. [8]

    How it works

    Security teams call this the patch window: the time between a fix and attackers using the flaw.

    Fortinet says AI shortens the time between a flaw being made public and attacks on it. [8]

    Seen it before

    A shop that learns a lock is weak changes it before word gets round.

    Fortinet used frontier AI to find nearly 100 flaws in its own products, and says AI also speeds up attacks once a flaw is public.

    The idea, in story 01: Why rules for hacking tools check the buyer, not the tool →
14. 15

    Ivanti and Citrix fix critical flaws

    What happened

    • Ivanti fixed 13 flaws in 5 products, and Citrix fixed 6 in XenServer, which runs virtual machines. [25]
    • The worst Ivanti flaw, rated 9.8 out of 10, lets an outsider run code with no login. [25]
    • Another Ivanti flaw, rated 9.4, lets outsiders add or delete the certificates a system trusts. [25]
    • Citrix's worst bug could let attackers work out a server pool's secret and take over the host. [25]
    • Ivanti says it has no evidence any of these flaws were used in attacks. [25]

    Who is involved

    It sits in Endpoint Manager, software firms use to manage their staff's computers. [25]

    Why it matters

    Companies on Endpoint Manager 2024 need update SU8 to close the worst flaw. [25]

15. 16

    Zoom and Asus fix serious flaws

    What happened

    • Zoom fixed a critical flaw in Zoom Workplace for Windows, rated 9.6 out of 10. [23]
    • It lets an outsider with no login gain more control of the computer than they should have. [23]
    • Version 7.1.8 and later carry the fixes for it and a second serious flaw. [23]
    • Asus fixed 4 flaws in some router software, which could let attackers run commands with full control. [23]
    • Zoom does not mention any attacks using its flaws. [23]

    Why it matters

    The fixes came among hundreds this week from firms including Apple, Intel, AMD and SAP. [23]

16. 17

    Siemens flags factory flaws, some never fixed

    What happened

    • Siemens published 9 notices covering 10 flaws in its factory and plant software. [12]
    • Siemens says fixes are not ready for several products, and some versions will never get one. [12]
    • It tells those customers to use its suggested workarounds instead. [12]
    • Two notices cover Node2Shell, a flaw attackers are already trying, though not yet on Siemens products. [12]
    • Germany's VDE CERT, a warning team for industrial firms, published 8 notices too, for makers including Beckhoff. [12]

    Why it matters

    Plants using Simatic SCADA and PCS 7, which run industrial processes, are among those affected. [12]

    The truth

    • Who pays Plants running Siemens versions that will never be fixed — They must keep running workarounds instead of a fix for as long as they use the equipment. [12]
17. 18

    Cyber incidents at a third of industrial firms

    What happened

    • A third of firms running factory and utility control systems had a cyber incident in the past year, SANS found. [26]
    • That is 34% of those asked, up from 22% last year, the biggest one-year jump SANS has seen. [26]
    • Incidents that hurt safety rose from 8% to 13% of firms asked. [26]
    • Only 15% say they could keep critical work going by hand for more than 48 hours. [26]

    Who is involved

    SANS, a security training group, asked nearly 700 IT and security staff. [26]

    Why it matters

    72% of firms said their incidents led to a facility or system being down. [26]

18. 19

    Microsoft fixes problems its own updates caused

    What happened

    • Microsoft's November updates fix Outlook sign-in and email problems that its October updates caused on Exchange servers. [27]
    • The October updates set a new memory limit on a Windows service that Exchange relies on. [27]
    • Some users could not sign in to Outlook on the web, and mail stayed stuck in the Outbox. [27]
    • The November updates also fix USB speakers and headsets that stopped working after September's security updates. [28]

    How it unfolded

    1. September USB audio breaks
    2. November fixed

    Why it matters

    Companies running their own Exchange email servers were the ones hit by the sign-in failures. [27]

    How it works

    Software engineers call this a regression: a change that breaks something that used to work.

    The limit hit Exchange servers by how busy they were, so some failed and some may never fail. [27]

    Seen it before

    Fixing a squeaky door hinge, then finding the door no longer shuts.

19. 20

    Glasswing's AI finds flaws, few get fixed

    What happened

    • VulnCheck, a flaw-tracking firm, says Anthropic's Project Glasswing has led to only 217 fixes in 7 months. [13]
    • Anthropic's dashboard counts 37,966 flaws found. [13] Its public list holds 7,349 of them. [13]
    • Daniel Stenberg, who runs cURL, found 19 of 34 Glasswing reports he checked were not flaws at all. [13]
    • Anthropic says it is "bottlenecked on our human capacity to validate these findings". [13]

    Who is involved

    Glasswing gives security researchers Anthropic's AI to find and fix flaws in software. [13]

    Why it matters

    Patrick Garrity of VulnCheck says the work is shifting onto the open-source maintainers Glasswing promised to help. [13]

    How it works

    Researchers call this triage: sorting reports to find the real ones first.

    Anthropic says it found more than 29,000 possible flaws but had reviewed only about 6,000 by 8 October. [13]

    Seen it before

    A teacher handed 100 essays a night marks the first few and the rest wait.

    The truth

    • Anthropic did not answer Infosecurity's request for comment. [13]
20. 21

    OpenAI builds a kill switch for its models

    What happened

    • OpenAI built a kill switch, called Control Room, to stop its research AI models quickly. [29]
    • It can stop a model's output 35 seconds after the decision to pull the plug. [29]
    • The delay comes from a store of saved permission checks, and OpenAI wants to cut it to zero. [29]

    How it unfolded

    The Register suggests recent cases of AI agents going beyond their task, like attacking Hugging Face, may explain the timing. [29]

    Why it matters

    OpenAI says being able to stop models matters more as they grow more powerful. [29]

    How it works

    Engineers call this fail-safe design: when the safety system breaks, everything stops instead of running on.

    If Control Room itself goes down, model output stops, OpenAI says. [29]

    Seen it before

    A lift's brakes grip when its power cuts, so a failure stops the lift.

21. 22

    Finland and Estonia want an EU AI emergency plan

    What happened

    • Finland and Estonia want the EU to have a plan for responding to serious AI incidents. [30]
    • They wrote ahead of Friday's first-ever meeting of EU leaders devoted to AI. [30]
    • Concern grew after AI agents went rogue this summer and hacked into other companies, Politico reports. [30]
    • Hungary has proposed national contact points for when an AI maker loses control of a model. [30]
    • It also wants a team of pre-vetted, security-cleared experts ready before incidents happen. [30]

    Who is involved

    • Security-cleared experts ready before incidents

    Why it matters

    The letter says Europe needs its own capacity to test the most advanced models for safety risks. [30]

22. 23

    CISA sets its priorities after staff cuts

    What happened

    • Nick Andersen, acting head of CISA, the US cyber-defence agency, set out its priorities in a 2 November memo. [31]
    • It tells staff to focus on infrastructure where a disruption would hurt the country most. [31]
    • The memo lists 13 needs the agency must meet to stay ready, several about staffing. [31]

    How it unfolded

    CISA has lost staff to cuts and is nearing its third year without a Senate-confirmed director. [31]

    On Monday, President Trump nominated a Palantir executive to lead it. [31]

    Why it matters

    The memo says CISA should give its help first to the most systemically important organisations. [31]

23. 24

    Three security start-ups raise money

    What happened

    • Bay, a Tel Aviv start-up, came out of hiding with $50 million to control what companies' AI agents may do. [32]
    • Its tool checks every action an agent proposes and can allow, redirect or block it within milliseconds. [32]
    • Dreeze raised $33 million to rebuild whole cloud apps after an attack, not just restore their data. [33]
    • Finite State raised $27 million to check the code that ships inside cars, medical devices and factory gear. [34]
    • Finite State ties its round to a wave of AI-found flaws in October. [34]

    Why it matters

    Finite State's customers face rules like the EU Cyber Resilience Act, which demand proof of product security. [34]

    The truth

    • Who gains Security start-ups Bay, Dreeze and Finite State — Each tied its pitch to AI agents or AI-found flaws, and each raised money this week. [32][33][34]
24. 25

    AI makes online abuse of Arab women worse

    What happened

    • Women in the Middle East seem to face more online abuse than average worldwide, and AI worsens it, DW reports. [35]
    • In a 2025 study, it found 49% of women online there do not feel safe from harassment. [35]
    • For activists, journalists and politicians, the figure rises to almost 70%. [35]
    • About a quarter of women abused online at work had met abuse made or spread by AI tools. [35]
    • Only about 7% of women facing such abuse report it to police, other studies suggest. [35]

    Who is involved

    UN Women, the UN body for women's equality, says 68% of women in the region have faced it. [35]

    Why it matters

    UN Women found 12% of women who faced online abuse were physically punished by their families for it. [35]