Cybersecurity · Saturday, 31 October 2026

A Chinese state group used an AI attack tool to reach about 140 targets in 11 days, CrowdStrike says

26 stories 33 sources

Signals on this beat

The bar and the word are our reading of how hard each one is pushing, and the arrow is where it is heading.

AI doing attackers' work Building↑

A Chinese state group's AI tool ran over 1,500 commands an hour against 200 targets. [1]
An open Chinese AI model, GLM-5.3, finished 3 of 13 test attacks. [19]
Sophos says a ransomware gang let AI agents do much of its break-in work. [9]

Flaws attacked soon after the fix High↑

Defiant saw the first attack on a WordPress flaw nine minutes after its guard rule went live. [12]
Drupal says a flaw in its shop add-on is already under attack. [13]
US agencies must fix two Ivanti flaws by 2 and 13 November. [10]

Poisoned copies from trusted shops Building↑

A poisoned Claude Code add-on came from Cursor's own store, and up to 5,000 users got it. [5]
Trezor warns a wallet sold by a reseller may hide a spy chip. [23]

Personal records taken in bulk High→

Florida says Social Security numbers of 4.9 million aid applicants were exposed. [7]
Revolve says saved card numbers were taken, from 127,200 people in Texas alone. [22]
Character.AI's users' chats could be sold in its bankruptcy. [21]

The stories

26 stories 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.

01
A Chinese state group used an AI attack tool to reach about 140 targets in 11 days, CrowdStrike says

about 140

targets reached, or probably reached, in 11 days

out of more than 200 the group went after in at least 14 countries [1]

1,500 vs 40

commands an hour with ARTEX, against a person typing

40 an hour is CrowdStrike's figure for hands-on attackers [1]

126

AI-driven attacks CrowdStrike counted in October so far

against about 45 in September [1]
What happened
- A Chinese state-backed group got into, or probably got into, about 140 targets in 11 days. [1]
- CrowdStrike, a US security company, published its report on Thursday. [1]
- The group went after more than 200 organisations in at least 14 countries. [1]
- It used ARTEX, a program built to test defences that runs attack steps on its own. [1]
- Its targets included foreign ministries in Latin America, Europe, Asia and the Middle East. [2]
- In one European country it targeted the presidency and the prime minister's office. [1]
- It also went after 19 US organisations working in drugs, chips, defence, nuclear and finance. [1]
- Phone and internet companies in the US, Belgium and the Philippines were targets too. [1]
- A person typing commands by hand manages up to 40 an hour, CrowdStrike has found. [1]
- With ARTEX, the group passed 1,500 commands an hour. [1]
- Its requests came as often as every 10 seconds. [1]
- ARTEX splits the work between planning programs and worker programs that talk to each other. [2]
- A human operator steers them, for example toward finding administrator passwords. [2]
- The agents tried several ways in at once: websites, cloud services, mail and VPNs. [2]
- A VPN is the private tunnel staff use to reach their office network. [2]
- When one way in failed, the agents switched to another. [2]
- In one break-in, the agents used login keys nobody had changed since 2022. [2]
- With those keys they forged an administrator pass and read staff mailboxes. [2]
- In another, they used a service that let anyone in without a login. [2]
- Meyers says spreading attacks across 200 targets costs the group very little. [1]
- CrowdStrike counted about 45 AI-driven attacks by groups in September, and 126 in October so far. [1]
Person typing40 commands an hour
With ARTEX1,500 commands an hour
How many commands an attacker sends in an hour, by CrowdStrike's count
Who is involved

CrowdStrike calls the group Vapor Panda, and it has tracked it since February 2021. [2]
The group paired ARTEX with DeepSeek, a Chinese AI model. [1]
Adam Meyers leads CrowdStrike's work tracking attackers. [1]
- Vapor Panda a Chinese state-backed hacking group, as named by CrowdStrike; it ran the campaign
- US
- Belgium
- the Philippines
- Adam Meyers a CrowdStrike senior vice-president who leads its work tracking attackers
- CrowdStrike a US security company; it tracked the campaign and published the report
- DeepSeek a Chinese AI model; the group paired it with ARTEX
How it unfolded

Earlier this month Microsoft tied a different ARTEX campaign to another Chinese group. [1]
1. Feb 2021 CrowdStrike first spots the group [2]
2. Early Oct Microsoft ties another ARTEX campaign to a different Chinese group [1]
3. 17 Oct the campaign begins [1]
4. Thu 29 Oct CrowdStrike publishes its count [1]
Why it matters

Meyers says the methods are old; only the speed and scale are new. [1]
He says defenders need to see what happens on their computers and logins. [1]
How it works

Security engineers call all the ways into a system its 'attack surface'. [2]
A tool that tries many ways at once finds whichever one was left open. [2]
Every feature left switched on is another way in

Security engineers call it shrinking the attack surface: a part that is switched off cannot be attacked, even before anyone knows its flaw.

The twist

A flaw in a feature that is switched off cannot hurt you. So turning things off protects you even before anyone knows the flaw is there.

The picture

Block Sites switched onexposed
Block Sites switched offsafe
One of Netgear's router flaws works only when the Block Sites feature is on. The same router with it off is not exposed to that flaw.
Step by step
1. Every feature a system runs is code strangers can reach
2. Each piece of code may hide a flaw nobody has found yet
3. An attacker needs only one of those flaws
4. A feature that is switched off cannot be used, flaw or no flaw
5. So switching off what is not needed removes ways in before any fix exists
The same force, elsewhere today

Where this chain is also running, in today's other stories.
- The poisoned Claude Code add-on
  
  The backdoor opened only on computers where SSH, a remote-login service, was already running; where it was off, the door stayed shut.
- Netgear's router fixes
  
  One flaw works only when the Block Sites feature is switched on, so routers with it off were never exposed.
- The shared ChatGPT link
  
  The way in was a second feature, a maths keyboard, loaded into the page and listening to anyone.
- The AI attack campaign
  
  The agents tried websites, cloud services, mail and VPNs at once, and in one break-in walked through a service that asked for no login.
Where you've seen this

A house

every window and door is a way in, and a bricked-up window cannot be forced open

A phone

Bluetooth and Wi-Fi left on in public are ways in that a switched-off radio does not offer

A shop at night

staff lock the side doors and leave one entrance they can watch

The catch

Switching a feature off also takes away what it does, so people switch it back on. And an attacker who is already inside can switch things on, which the Cursor malware tried to do with SSH.

And the whole of it

Every phone, router and work system arrives with many parts switched on by its maker. Most people who use them never learn which parts are running, and an attacker only needs one of them.
The truth
- What is really going on CrowdStrike says a Chinese state-backed group used ARTEX, an AI attack tool, to go after more than 200 organisations in 11 days, at over 1,500 commands an hour. [1]
- What is really going on The tool did not need new tricks: in one break-in it used login keys nobody had changed since 2022. [2]
- Why it works on people Talk of AI attackers makes the danger sound brand new, but CrowdStrike's own Meyers says only the speed and scale are new. [1]
- The count of 140 includes targets CrowdStrike only judges likely, not all confirmed. [1]
- A second report on the same research counts confirmed follow-up activity at 9 organisations. [2]
- Who gains Vapor Panda, the Chinese state-backed group — An AI tool lets a small team chase far more targets for very little cost, Meyers says. [1]
- Who pays Foreign ministries, telecoms and 19 US organisations — They were among Vapor Panda's targets, and CrowdStrike says the campaign is still running. [1]
- Nobody knows yet
  How many of Vapor Panda's targets were really broken into.
  
  CrowdStrike told CyberScoop about 70% of targets were confirmed or judged likely compromised, about 140. [1] The Hacker News reports confirmed follow-up activity at nine organisations. [2] The two counts measure different things.
- Nobody knows yet
  Which European government and which foreign ministries were hit.
  
  CrowdStrike lists them by region and type, not by name. [1][2]
What comes next

CrowdStrike says the campaign is still running. [1]

02
Poisoned Claude Code add-on in Cursor's store
What happened
- A poisoned copy of Claude Code sat in Cursor's add-on store for about 11 hours. [3]
- Cursor says at most 5,000 users got the poisoned copy. [5]
- It tried to plant a backdoor, a hidden way back in, through SSH. [3]
- SSH is a tool for logging in to a computer from far away. [3]
- It also sent each computer's user name, internet address and Git email to the attacker. [3]
- Cursor shut its whole store to check every add-on, then reopened it. [4]
Who is involved

Claude Code is an AI helper for writing software, made by Anthropic. [3]
Cursor is an app for writing code, owned by SpaceX. [5]
- Claude Code made by Anthropic
- Cursor owned by SpaceX
Why it matters

Add-ons in Cursor update themselves by default, so some users got it without choosing to. [4]

How it works

Mistakes in the code meant the backdoor worked only where SSH was already running. [3]

The backdoor opened only on computers where SSH, a remote-login service, was already running; where it was off, the door stayed shut.
The idea, in story 01: Every feature left switched on is another way in →
The truth
- Anthropic says: 'it wasn't a compromise of Anthropic's systems or employee laptops.' [6]
- Neither Cursor nor Anthropic has said how the fake reached the store. [4]
- Who pays Up to 5,000 developers who got the poisoned add-on — They must check for the attacker's key and, where hit, rebuild their machines and change their secrets. [3]
- Nobody knows yet
  How the poisoned Claude Code add-on got into Cursor's store under Anthropic's name.
  
  Anthropic says its systems were not broken into, and Cursor has not explained what happened. [6][4]
Good news
- Cursor worked with the host to take the attacker's server offline. [3]
- Mistakes in the poisoned Claude Code add-on kept its backdoor shut on most of the up to 5,000 machines it reached, and Cursor had the attacker's server taken offline.
03
Florida aid site leaks 4.9 million people
What happened
- Florida says a break-in exposed data on 4.9 million people who applied for student aid. [7]
- The data includes Social Security numbers, the US ID number used for tax and credit. [7]
- It also includes names, addresses, phone numbers and dates of birth. [7]
- It took the aid website down and told the state attorney general. [7]
- It says student aid payments and tuition will not be delayed. [7]
How it unfolded

Florida's education department learned of the break-in on 19 October. [7]
A week earlier, California's education department reported its own pupil-records breach. [7]
It is the state's second big government breach this year, after its motor vehicle agency. [7]
1. 19 Oct break-in found
2. Thursday 4.9 million told
3. A week earlier California breach
4. This week Florida breach
Why it matters

Those affected are being told to watch their credit reports and bank accounts. [7]
The truth
- The department will not say who got in or how. [7]
- Who pays 4.9 million Florida student aid applicants — Their Social Security numbers and dates of birth were exposed, and they are told to watch their credit reports. [7]
- Nobody knows yet
  Who broke into Florida's student aid site, and how.
  
  Florida's education department declined to comment on the attackers or the method. [7]
04
Apple warns spyware targets in 99 countries
What happened
- Apple sent new warnings to people it thinks were hit by paid-for spyware. [8]
- This round went to users in 99 countries, Apple said. [8]
- Spyware is software that secretly watches what a person does on their phone. [8]
- Those warned include journalists and documentary filmmakers. [8]
- Apple does not say who it thinks is behind the attacks. [8]
Who is involved
- Apple sent the warnings
- Journalists
- Filmmakers
- Citizen Lab
How it unfolded

Since it started, Apple has warned users in over 150 countries. [8]

Why it matters

Apple and the Citizen Lab, a Toronto research group, advise turning on Lockdown Mode. [8]
The truth
- Access Now, a rights group, says 'repressive states continue to use these powerful surveillance technologies'. [8]
What comes next

Talks on rules for spyware sellers, called the Pall Mall Process, will soon reach final rounds. [8]
05
Iran suspected over spoiled food at US bases
What happened
- US officials suspect Iran-linked hackers made freezers fail at about 6 military base shops. [9]
- The failures spoiled food, starting in late August, Nextgov, a US news site, reports. [9]
- At Fort Huachuca in Arizona, a freezer switched to defrost and warmed its contents. [9]
- US Cyber Command and Defense Department police are investigating. [9]
- They are working with France, where hackers overrode cooling at some 350 places this month. [9]
Who is involved
- US Cyber Command
- Defense Department police
How it unfolded
1. Late August base freezers start failing
2. This month French cooling overridden
Why it matters

If officials are right, a foreign state reached the food stores on US bases. [9]
The truth
- No government has publicly named who did it. [9]
- Nobody knows yet
  Who made the freezers fail at US military base shops.
  
  Officials suspect Iran-linked hackers, but that comes from unnamed people speaking to Nextgov, not an official statement. [9]
06
Last Ivanti flaws fixed after attacks
What happened
- CISA, the US cyber-defence agency, says attackers are using two more Ivanti flaws. [11]
- Attackers joined three flaws to take full control of these boxes without a password. [10]
- Ivanti now has fixes for all 3. [10]
Who is involved

Ivanti makes Connect Secure, a box staff use to reach work networks from home. [10]
- CISA US cyber-defence agency
- Ivanti makes work VPN boxes
How it unfolded

CISA listed the first of the three on 13 October, and the other two on Friday. [10]
1. 13 Oct first flaw listed
2. Friday two more listed
3. 2 Nov first deadline
4. 13 Nov second deadline
Why it matters

The attackers searched for administrator passwords and BitLocker keys, which unlock scrambled disks. [10]

How it works

Security engineers call this an exploit chain: small flaws joined into one big attack. [10]
The truth
- Ivanti shipped its main fix first, so the other two stayed open for over two weeks. [10]
What comes next

US federal agencies must install one fix by 2 November and the other by 13 November. [10]
07
WordPress flaw attacked minutes after the fix
What happened
- Attackers went after a new WordPress flaw within hours of its fix, Defiant says. [12]
- The flaw is in how WordPress unpacks its own updates. [12]
- For seconds, new files sit in a web-reachable folder on many servers. [12]
- Defiant saw the first attack 9 minutes after its guard rule went live. [12]
- It counted more than 18,500 attempts in the first 21 hours. [12]
- WordPress fixed it in version 7.1.4, with 22 security fixes in all. [12]
Who is involved

WordPress is free software for building websites. [12]
Defiant is a web security firm behind Wordfence, a guard for WordPress sites. [12]
- WordPress website software
- Defiant makes Wordfence
How it unfolded
1. Rule goes live Defiant's guard
2. 9 minutes later first attack
Why it matters

So sites that wait even a day to update face the attempts. [12]

How it works

Engineers call this a race condition: the attack works only inside a short time gap. [12]
Good news
- Defiant has seen no successful break-ins yet. [12]
08
Drupal shop add-on attacked
What happened
- Drupal warns that attackers are using a flaw in Commerce Core, its online-shop add-on. [13]
- More than 42,000 sites use Commerce Core, including research universities. [13]
- The flaw lets strangers reach parts of a shop they should not, without logging in. [14]
- Drupal rates it 21 out of 25 for danger. [14]
- Shop owners are told to update and check their orders for anything odd. [14]
Who is involved

Drupal is free software for building and running websites. [13]
- Drupal website software
Why it matters

Drupal warns the flaw 'may financially impact the vendor', meaning the shop owner. [13]

How it works

Only sites running the shop add-on are exposed to this flaw. [13]
The truth
- Drupal has not said who is attacking or how many shops were hit. [14]
09
Second firm blames North Korea for Binance theft
What happened
- Chainalysis, a firm that traces crypto, now also blames North Korea for the Binance theft. [15]
- Both firms name the Lazarus Group, also called TraderTraitor, with high confidence. [15]
- The thief is swapping coins that can be frozen into ether, which cannot, Decrypt reports. [16]
- More than $2 billion of those freezable coins has already been frozen. [16]
Who is involved

Binance's co-chief, Richard Teng, says it was not a one-click attack. [16]
- Chainalysis
- TRM Labs
- Richard Teng Binance co-chief
How it unfolded

Binance, the biggest crypto exchange, lost about $6.3 billion this week. [15]

Why it matters

Experts say North Korea's government funds itself through cybercrime, which it denies. [15]

How it works

Stablecoins have a company behind them that can freeze them; ether has none. [16]
The truth
- Every dollar swapped into ether before a freeze is out of the freezers' reach, Decrypt says. [16]
- Who gains Whoever stole from Binance — Coins swapped into ether before a freeze cannot be blocked by any company, Decrypt reports. [16]
What comes next

He promises a full technical account once every fact is checked. [16]
10
GCHQ lost track of thousands of drives
What happened
- GCHQ, Britain's spy agency for phones and the internet, cannot account for thousands of storage devices. [17]
- They include USB sticks, hard drives, discs and memory cards. [17]
- The Investigatory Powers Commissioner, the watchdog for spying powers, reported it on Thursday. [17]
- Some devices may hold data gathered under Britain's spying law. [17]
- GCHQ says its old tracking was not robust enough and it has acted on every recommendation. [17]
Who is involved
- Investigatory Powers Commissioner spying watchdog
How it unfolded

Commissioner Brian Leveson says the problems were raised with him several years ago. [18]
1. Years ago problem raised
2. Thursday still no count
Why it matters

The watchdog says it is 'very concerned' and unhappy at how long GCHQ's search has taken. [17]
The truth
- The report also says GCHQ errors led to 5 port stops of innocent people. [17]
- Who pays Innocent people wrongly flagged by GCHQ — Its errors led to 5 port stops where people were held and had their devices copied. [17]
- Nobody knows yet
  How many storage devices GCHQ has lost, and what is on them.
  
  The watchdog says GCHQ still has no clear picture of the scale, and GCHQ would not say whether anyone unauthorised has used them. [17][18]
What comes next

It plans an automatic system to track drives by the end of the year. [18]
11
Chinese AI model finishes test attacks
What happened
- GLM-5.3, an AI model from China's Z.AI, carried out complex attacks in controlled tests. [19]
- It finished 3 of 13 multi-step attack tests at least once, Irregular found. [19]
- It did not refuse any of the attack tasks it was given. [19]
- OpenAI's GPT-6 Astra finished 9 of 10 in a different test. [19]
Who is involved

Irregular is a research firm that tests AI safety. [19]
- GLM-5.3 made by Z.AI
- Irregular AI safety research firm
Why it matters

Anyone can download GLM-5.3 and change it, unlike top US models. [19]

How it works

Researchers call one habit it showed 'reward hacking': reaching the answer by breaking the rules. [19]
The truth
- The two tests were not the same, so the scores do not compare directly. [19]
12
A shared ChatGPT link could read your chats
What happened
- Researchers at Parameter showed a shared ChatGPT link could read the opener's private chats. [20]
- The victim only had to open the link and click Preview once. [20]
- Preview turns code in a chat into a small web page. [20]
- The way in came from a second feature: a maths tool loaded into the page. [20]
- It accepted messages from any sender, which let the attacker's code out. [20]
- They reported it on 18 October, and OpenAI fixed it 8 days later. [20]
Who is involved
- Parameter security firm
How it unfolded
1. 18 Oct reported
2. 8 days later fixed
Why it matters

Shared chat links travel by email, chat and social posts, so many people open them. [20]

How it works

A message between parts of a page is safe only if the receiver checks the sender. [20]

The way in was a second feature, a maths keyboard, loaded into the page and listening to anyone.
The idea, in story 01: Every feature left switched on is another way in →
Good news
- It was a controlled test, and no real attacks have been reported. [20]
- OpenAI fixed the shared ChatGPT link flaw eight days after Parameter's researchers reported it, and no real attacks using it have been reported.
13
US asks a court to guard Character.AI chats
What happened
- The US Justice Department's bankruptcy office wants Character.AI users' chats protected in its sale. [21]
- It has 12 million users a month, who spend about two hours a day on it. [21]
- Its privacy policy lets user data pass to a buyer if the company is sold. [21]
- The lawyers want a privacy official to advise the court on what can be sold. [21]
- Asked if chats are shared, the chatbot answers: 'Not without your explicit permission.' [21]
Who is involved

Character.AI is an app where people chat with made-up AI characters. [21]
- US Justice Department its bankruptcy office
- Character.AI chatbot app
How it unfolded

That policy was changed on 1 July, while the company was trying to sell itself. [21]
1. 1 July policy changed
2. Now sale in court
Why it matters

The filing asks for special care over children's data and health data. [21]
The truth
- A buyer could get users' most private chats under a policy changed during the sale. [21]
- Who gains A buyer of Character.AI — Its privacy policy, changed on 1 July during the sale, lets user data pass to a buyer. [21]
14
Revolve says card numbers were stolen
What happened
- Revolve, an online fashion shop, says hackers took customers' saved card numbers. [22]
- It told the SEC, the US markets regulator, in a filing on Thursday. [22]
- The thieves took names, contact details, card numbers and expiry dates. [22]
- It told Texas that 127,200 people there were affected, and 1,328 in Vermont. [22]
- It says it told US federal police and hired a security firm. [22]
Who is involved
- Revolve online fashion shop
- US federal police
- A security firm
- ASOS
- M&S
- Harrods
Why it matters

Fashion shops are a regular target: ASOS, M&S and Harrods were all hit in two years. [22]

How it works

Only customers who saved their card in their account were hit. [22]
The truth
- Revolve has not said how many people were hit in total. [22]
- Who pays Revolve shoppers who saved a card — Card numbers and expiry dates were taken, including from 127,200 people in Texas. [22]
- Nobody knows yet
  How many Revolve customers had card details taken.
  
  Revolve gave no total; only state counts are public, such as 127,200 in Texas. [22]
15
Trezor checks a wallet for a spy chip
What happened
- Trezor, which makes devices that hold crypto keys, warns one wallet may hide a spy part. [23]
- A buyer in Indonesia, who bought it in August, found an extra circuit board inside. [23]
- The buyer says it can read the recovery words, which rebuild a wallet, off the screen. [23]
- Trezor stopped selling through CryptoBilis on 9 October. [23]
Who is involved
- Trezor makes crypto wallets
- CryptoBilis reseller
How it unfolded

It came from CryptoBilis, a reseller already linked to tampered wallets from Ledger, a rival maker. [23]
1. August the wallet is bought
2. 9 Oct Trezor stops CryptoBilis sales
Why it matters

Ledger users have already lost money; Singapore police added $3,500,000 to the total. [23]
The truth
- It has not yet received the device to examine it. [23]
- Nobody knows yet
  Whether the reported Trezor wallet really holds a spy chip.
  
  Trezor has not yet received the device to examine it. [23]
Good news
- Trezor says it knows of no money stolen this way. [23]
16
Defence contractor admits helping North Korean job scheme
What happened
- John Allen Kurrle Jr., 62, a US defence contractor, pleaded guilty to helping North Korean workers. [24]
- He let them use his name and top secret clearance to apply for jobs. [24]
- The scheme sent at least 88 applications and won 9 job offers. [24]
- One company noticed the worker on a video call did not match Kurrle's age. [24]
Who is involved
- John Allen Kurrle Jr. defence contractor, 62
How it unfolded

North Korea uses borrowed identities to get remote jobs and earn money for its military. [24]

Why it matters

North Korea targets contractor jobs because they are the easiest way to sensitive work. [24]
The truth
- He says he did not know North Korea was involved, but does not dispute it benefited. [24]
What comes next

He faces up to 20 years and will be sentenced on 28 January 2027. [24]
17
Syrian hacker behind fake AP tweet jailed
What happened
- Ahmad Umar Agha, a Syrian hacker, was sentenced to 42 months in a US prison. [25]
- In 2013 he posted a false report of Obama being hurt on the AP's Twitter account. [26]
- The hoax briefly wiped $136 billion off US share prices. [26]
- He hacked about 47 media groups, charities and government bodies, prosecutors say. [26]
- The US first charged him in 2014; he was taken into custody last year. [25]
Who is involved

He worked for the Syrian Electronic Army, run by people close to Bashar al-Assad. [26]
- Syrian Electronic Army served Assad's circle
How it unfolded

Assad was Syria's president until rebels overthrew him in 2024. [25]
1. 2013 fake tweet
2. 2014 charged
3. This week sentenced
Why it matters

Prosecutors say his hacks tried to silence reporting on the Assad government's atrocities. [26]
The truth
- Firas Dardar, a second member, is still on the run. [25]
Good news
- Ahmad Umar Agha, who used the AP's Twitter account in 2013 to post a fake report of an attack on Barack Obama, was sentenced to 42 months in prison.
18
TeamViewer flaw lets a host attack the helper
What happened
- A TeamViewer flaw lets a computer being helped take over the helper's computer. [27]
- Remedio, a security firm, found that a fake screen size overflows the helper's memory. [27]
- It is rated 8.8 out of 10 for danger. [27]
- Version 15.83 of TeamViewer fixes it, along with several other flaws. [27]
Who is involved

TeamViewer is software that lets one person control another's computer from far away. [27]
- TeamViewer remote-control software
Why it matters

The people at risk are support technicians, who connect to other people's computers. [27]

How it works

Programmers call this an integer overflow: a number too big for its space wraps round. [27]
Good news
- TeamViewer says it knows of no attacks using these flaws. [27]
19
Attack on Alpine Linux's code site failed
What happened
- Attackers tried a GitLab flaw on Alpine Linux's own code site on 20 and 21 October. [28]
- CISA says the GitLab flaw is now being used in attacks. [28]
- The attack got partway 4 times and then failed, Alpine says. [28]
- Alpine rebuilt the site from scratch and reset every password. [28]
Who is involved

Alpine Linux is a free version of the Linux operating system. [28]
GitLab is software that teams use to store and build code. [28]
- Alpine Linux free operating system
- GitLab code storage software
How it unfolded
1. 20 Oct attack starts
2. 21 Oct attack ends
Why it matters

Updating GitLab alone does not fix sites that run their own copy of Ruby. [28]
The truth
- Alpine found no sign the attacker ran code, but says it cannot rule it out. [28]
Good news
- Its software packages and its 21 October releases were not tampered with. [28]
- Attackers who went after Alpine Linux's code site failed, and its software packages were not tampered with.
20
Flood attacks up 38% in three months
What happened
- Cloudflare blocked 18.2 million flood attacks from July to September, it said on Thursday. [29]
- That is 38% more than in the three months before. [29]
- A flood attack, or DDoS, swamps a site with junk traffic until it stops answering. [29]
Who is involved

Cloudflare is a company that carries and protects a large share of web traffic. [29]
- Cloudflare web traffic company
- Dysphoria
- Evooo1Bot
How it unfolded

It counted 13.2 million from April to June and 10.0 million from January to March. [29]

Why it matters

The rise was linked to attack services that anyone can hire. [29]
The truth
- New attack networks, called Dysphoria and Evooo1Bot, have appeared since, Cloudflare warns. [29]
Good news
- The very biggest attacks fell 69%, likely after police action against attack networks in March. [29]
- Cloudflare counted 251 of the very biggest flood attacks on websites from July to September, down from 805 in the three months before.
21
UK to look at making software makers pay for breaches
What happened
- The UK government will consult next year on making tech firms liable for insecure products. [30]
- Liable means a company can be made to pay for harm its product causes. [30]
- The pledge was in the budget documents published on Wednesday. [30]
How it unfolded

Until now Britain has relied on voluntary codes for software makers. [30]
1. Wednesday budget pledge
2. Next year consultation
Why it matters

The 2025 hack of Jaguar Land Rover cost Britain's economy an estimated $2.5 billion. [30]
The truth
- The government has not said who will lead it or when it will start. [30]
- Who gains Software makers selling into Britain — They still face only voluntary codes until a consultation that starts next year at the earliest. [30]
22
Netgear fixes flaws in home routers
What happened
- Netgear has released fixes for flaws in several of its home routers and Wi-Fi extenders. [31]
- An attacker must already be on the home network to use them. [31]
- Block Sites lets owners stop the router reaching chosen websites. [31]
- Several models have reached end of service, so they will get no more updates. [31]
Who is involved
- Netgear router maker
Why it matters

Netgear advises owners of the oldest models to buy a newer router. [31]

How it works

One flaw works only if the router's Block Sites feature is switched on. [31]

One flaw works only when the Block Sites feature is switched on, so routers with it off were never exposed.
The idea, in story 01: Every feature left switched on is another way in →
The truth
- For one Orbi model, even the newest software does not fix the flaw. [31]
23
Microsoft finally fixes broken work VPNs
What happened
- On Wednesday Microsoft fixed Always On VPN links that its September updates broke. [32]
- Always On VPN links a work laptop to the office network whenever it goes online. [32]
- After the update, some connections got stuck on 'Connecting' or failed. [32]
How it unfolded
1. September updates break VPNs
2. Wednesday fix confirmed
3. Now optional preview update
4. November regular monthly updates
Why it matters

The problem began in late September, so some staff went about a month without working VPNs. [32]
The truth
- Microsoft had wrongly said in October that the problem was already fixed. [32]
What comes next

The fix is in an optional preview update; others can wait for November's updates. [32]
24
Cloudflare AI service showed users others' data
What happened
- Cloudflare says its AI service sent 155 answers holding data from other customers' requests. [9]
- It happened between 27 August and 9 September, with one AI model. [9]
- Only very long requests were affected. [9]
- A setting mismatch and a bug in SGLang, software that runs AI models, caused it. [9]
How it unfolded
1. 27 Aug leak starts
2. 9 Sep leak ends
Why it matters

Leftover data from one customer's session could be read by a later long request. [9]

How it works

Engineers call that leftover memory a cache: kept to speed up answers, it held old data. [9]
Good news
- A researcher reported it, and Cloudflare found no sign anyone used it on purpose. [9]
25
A ransomware gang let AI agents do the work
What happened
- Sophos, a security firm, says a gang it calls Gold Ogre used AI heavily. [9]
- Ransomware locks a company's files until it pays. [9]
- AI agents run through Cursor did much of the break-in with little human help. [9]
- They mapped networks, guessed logins and moved from machine to machine. [9]
Who is involved
- Gold Ogre ransomware gang
Why it matters

SecurityWeek groups it with other reports of AI moving deeper into ransomware. [9]
The truth
- The gang got past AI safety rules by calling its attacks authorised tests. [9]
26
A free checklist for AI systems and ransomware
What happened
- A free online checklist helps companies test their AI systems against ransomware. [33]
- Behnaz Karimi built it over two years with Yuvaraj Govindarajulu. [33]
- Teams score their own defences by hand; it does not scan their systems. [33]
- Visa problems stopped Karimi presenting it at a Denver conference this week. [33]
Who is involved

It is called RAID-AI and comes from people at OWASP, an open security group. [33]
- RAID-AI free checklist
- OWASP open security group
Why it matters

Karimi says ransomware attacks on AI systems are now starting to happen. [33]

A Chinese state group used an AI attack tool to reach about 140 targets in 11 days, CrowdStrike says

26 stories on this beat today.

Every feature left switched on is another way in