Cybersecurity · Tuesday, 20 October 2026
One stolen login let an intruder copy the school records of millions of California pupils over four months
Someone used a stolen login to copy files on California's public school pupils from ETS, the state's testing company, from May to September; the records include disability and test details. Ahold Delhaize confirmed its outage was an attack, and one hacker's AI programs broke into more than 100 websites.
5.7 million+
pupils in California's public schools in 2025-26
the copied files cover everyone enrolled in the last three school years
4 months
that one stolen login was used, from 22 May to 18 September
ETS noticed on 25 September
2006
the earliest year of school-leavers whose records were taken
pupils whose move to another school was never confirmed to the state
The lead story — what happened
-
Someone copied files about California's public school pupils from a file-sharing site run by ETS, a non-profit that runs the state's school tests.
[1] [2] -
ETS used the site to send test data to the California Department of Education, the state body that runs public schools.
[1] -
The intruder signed in with one real login that had fallen into the wrong hands, again and again from 22 May to 18 September.
[1] -
ETS only spotted the activity on 25 September, a week after the last known visit.
[1] -
The files cover every pupil enrolled in the last three school years, and some who left as far back as 2006.
[1] [2] -
No count has been given, but California's public schools had more than 5.7 million pupils in 2025-26.
[1] -
The records hold names, birth dates, home addresses, schools, grades, ethnicity and first language.
[1] [2] -
Some also show a child's disability, the extra help they get in tests, and their test results.
[1] [2] -
The files held no Social Security numbers, passwords or bank details, and nothing about parents.
[1] -
ETS closed the account, took the site offline and is building a new one with stronger protection.
[1] -
Letters to families began on 16 October and will go out into early November, with two years of free identity protection, a service that warns people when their details are misused.
[1] -
Neither ETS nor the department has said how the login was stolen or who used it.
[1]
Who is involved
-
ETS (Educational Testing Service)
a US non-profit that runs school tests for states under contract; its file-sharing site was broken into
-
California Department of Education
the state body that runs public schools; it published the breach notice and is writing to families
-
California pupils and their families
millions of current and former pupils whose records were copied; they did not choose ETS
How it unfolded
-
22 May the stolen login is first used on ETS's file-sharing site
[1] -
18 Sep the last known use of the login
[1] -
25 Sep ETS spots the activity on the site
[1] -
2 Oct the state tells school district heads in a letter
[1] -
16 Oct letters to families begin, running into early November
[1]
Where this points
Watch whether the copied files appear online; ETS says its web monitoring has found no sign of that so far.
What is pushing on the whole day
The bar and the word are our reading of how hard each one is pushing today. The arrow is where it is heading. The evidence is in the stories below.
One stolen login gave an intruder four months inside ETS's file-sharing site.
OpenCode, an AI coding tool, stored its users' keys for paid AI services in one system, and attackers reached it.
One hacker's AI programs broke into more than 100 websites for about $50 each.
Moody's found flaws already used in attacks take a median of 87 days to fix.
The rest of the day
26 more stories on this beat.
Each with its own sources. None of these is a link to the story above.
-
02
Grocery giant confirms it was attacked
Ahold Delhaize, a Dutch company that owns supermarket chains in Europe and the US, said on Monday that 'malicious actions of others' caused the trouble it reported on 12 October.
[3] [4] It had taken websites, apps and other systems offline across most of its brands, including Stop & Shop and Food Lion in the US.[4] Its shops stayed open, but some struggled with supplies.[5] It reported the attack to Dutch police on Monday and says it has no evidence that customer data was taken.[3] [5] Why it matters — Online orders are about a tenth of sales at Albert Heijn, the biggest Dutch supermarket chain, and it will not deliver them before Thursday, at least 11 days after the outage began.
[5] Ahold Delhaize has not said what kind of attack it was or whether a ransom was demanded.[3] -
03
Backup-key attackers tried 1,000 firms
ConnectWise, which sells software that IT firms use to look after their clients' computers, still had its backup and management services switched off on Monday, with no date to restart.
[24] Huntress, a security company, now says the attackers' internet address tried to sign in to nearly 1,000 companies' Microsoft 365 accounts.[23] More than 300 of those sign-ins worked, though not all of them reached any data.[23] A second wave aimed at groups working on policy, national security, trade and vital minerals mostly failed.[24] Why it matters — ConnectWise says it cannot see inside its customers' systems, so each one has to check its own accounts.
[24] While the services are off, backups do not run and saved copies cannot be restored.[23] -
04
AI coding tool's stored keys exposed
OpenCode, a free AI tool that helps people write software, says attackers got into an old system it was in the middle of shutting down.
[7] The system held keys that some users had saved to pay for AI from OpenAI, Anthropic and Google.[7] About 1.4% of users of its Console service were hit, from 17 October.[6] OpenCode called it part of a 'distillation attack', where someone sends huge numbers of questions to an AI to copy its answers into a cheaper one.[6] Why it matters — Some users reported more than $1,700 of charges they did not make, run up overnight on their own keys.
[6] OpenCode has told the three AI firms, and says users who did not get its email were not affected.[6] [7] -
05
One hacker, AI programs, 100 websites
A researcher at NorForge, a security company, watched one hacker set AI programs loose on websites with fewer than 100 characters of instructions.
[14] The programs tried more than 1,200 sites and broke into more than 100, including sites linked to Deere, SharkNinja and Cummins.[14] [15] A typical break-in took about 23 hours of trial and error.[14] The researcher puts the cost at about $5,000 of AI use, or about $50 a break-in.[14] Sites tried1,200Sites broken into100About one in twelve attempts worked, at roughly $50 of AI use for each success. Why it matters — Deere says it has no sign that sensitive data was reached, and SharkNinja says only a product-help site was hit.
[14] The researcher could watch because the hackers left Cairn, a free AI tool they were using, open to the internet.[14] [15] -
06
Spyware found on a Serbian rector's phone
Citizen Lab, a University of Toronto lab that studies spyware, confirmed on Monday that Pegasus was on the phone of Vladan Dokic, head of the University of Belgrade.
[17] [18] Pegasus is spyware made by NSO Group, an Israeli company, and this infection needed no click at all.[18] Dokic backs the student protests against Aleksandar Vucic, Serbia's former president, who is running to be prime minister on 25 October.[17] A prosecutor in a corruption case against government officials, Rade Bajic, got an Apple spyware warning in August.[17] [18] Why it matters — SHARE Foundation, a Serbian digital-rights group, says at least 15 people in Serbia have been targeted with such spyware this year.
[17] NSO did not answer questions, and it has long said it does not work with governments that abuse its tools.[17] -
07
Judge will not punish NSO for WhatsApp
A US federal judge in California refused on Friday to hold NSO Group, the maker of Pegasus, in contempt of court.
[16] WhatsApp, Meta's messaging app, said NSO had broken a 2025 court order banning it from using WhatsApp to attack people.[16] It pointed to 23 new accounts it linked to NSO, but would not say exactly how it found them.[16] The judge said that left too little evidence to prove it.[16] Why it matters — WhatsApp said showing its methods would help spyware makers hide in future.
[16] Access Now, a digital-rights group, said forcing that choice would let spyware makers escape punishment.[16] -
08
Recruitment software switched off
Bullhorn, a US company whose software recruitment agencies use to track job candidates, sign up new staff and run payroll, switched its platform off after a cyberattack.
[8] Customers first had trouble posting jobs on Friday, and on Saturday it confirmed an attack.[8] It says it has more than 10,000 customers worldwide.[8] It has not yet found out whether data was taken and has given no date for the return of most services.[8] Why it matters — Job seekers' details pass through this software as well as the agencies' own.
[8] Users who do not sign in through their employer's system must reset their passwords.[8] -
09
Russian server host loses customer files
FirstVDS, one of Russia's oldest companies renting out servers, says hackers stole data from some of its customers' servers.
[22] They used a flaw in VMmanager, software from a firm called ISPsystem that runs many servers at once.[22] Their tool copied passwords, cloud logins, settings files and crypto wallets without changing anything, so the servers show no trace.[22] The flaw is not fully fixed yet.[22] Why it matters — FirstVDS told affected customers to assume their data is gone and to change every password.
[22] It has not said how many customers were hit or who did it.[22] -
10
Poisoned copy of a website add-on
Rank Math PRO is a paid add-on for WordPress, the software behind many websites, that helps sites show up in search results.
[10] For 11 hours on 15 and 16 October, some copies downloaded from Rank Math's own website carried hidden code that let attackers take over a site.[10] [11] The company removed the files and released a clean version, 3.0.124.[10] The free version, used on more than 4 million sites, was not affected.[10] Why it matters — Updating does not remove the hidden code, so owners who switched the bad copy on must restore the site from a backup or clean it by hand.
[10] [11] The code could also reach passwords and keys stored on the web server.[10] -
11
Home media servers attacked a day later
Plex Media Server is software people run at home to stream their own films and music.
[13] On 16 October a researcher published four flaws that, used together, let a stranger run programs on an exposed server without a password.[12] The next day another researcher's decoy server caught real attacks using that code.[12] Plex had already shipped fixes by 1 September.[12] Why it matters — Infected servers were set to mine cryptocurrency and to pass strangers' internet traffic through the owner's home connection.
[13] Criminals want home connections so they can reach company email from an ordinary-looking address, the second researcher wrote.[12] -
12
Nikkei's third breach in a month
Nikkei, the Japanese business news group that owns the Financial Times, has reported its third data breach in about a month.
[9] Attackers used stolen logins on 7 October to reach a system the company uses to build software.[9] It held 1.83 million records, of which about 150,000 had names, emails, addresses, phone numbers or birth dates.[9] Nikkei says it has no sign the data was misused.[9] Why it matters — The two earlier breaches also began with one staff account taken over, one in Google's office software and one in Microsoft's.
[9] People in the 150,000 records are to be told one by one.[9] -
13
US agencies told to drop DeepSeek
The US Department of Homeland Security has banned civilian government agencies from using DeepSeek, a Chinese AI company, and anything built with its models.
[19] [21] Agencies must also remove what they already run.[19] The order says data typed into DeepSeek is stored in China, where China's law lets its government demand it.[19] It does not cover the Department of Defense.[21] Why it matters — Services such as OpenRouter can pass a request to DeepSeek's models without the user picking it, so agencies may be using it without knowing, The Register notes.
[20] The reports differ on when the ban began: a letter dated 26 June, or the end of July.[19] [21] -
14
$80m deal over 62 million records
Conduent, a company that runs back-office work for US government health programmes and toll roads, agreed to pay $80 million to settle a lawsuit.
[33] The case is about a 2024 hack that exposed personal and health data of about 62 million people.[33] People with documented losses from fraud can claim up to $3,500 each.[33] A court must still approve the deal.[33] - Data in hard-to-read backup files40 m · 65%
- Data in files easy to read22 m · 35%
The 62 million people are split into two groups, by how easily the stolen files could be read. Why it matters — Conduent denies fault, and says it has cyber insurance and does not expect the deal to hurt its finances.
[33] It took almost three months to find the 2024 break-in.[33] -
15
Negotiator suspected of running extortion site
KrebsOnSecurity, a security news site, reports that the FBI believes an executive at a large US consulting firm helped run Shinywire.
[34] Shinywire was a website ShinyHunters, a hacking group, used to pressure nearly 100 victims into paying, from May until 16 October.[34] The FBI says the group took more than $70 million from victims this year.[34] Nobody at the firm has been charged, and Krebs does not name it.[34] Why it matters — Victims hire negotiators to deal with hackers for them, often on their insurer's advice.
[34] The FBI arrested Edward Dubrovsky, a Canadian negotiator, earlier this month in the same investigation.[34] -
16
Siemens tool used to slip in malware
Siemens, the German industrial company, said on Monday that attackers used a flaw in Siemens Software Center, its program for installing customers' software.
[26] They bundled a genuine, signed copy with their own files, which the program loaded without checking.[26] The attacks are part of Romulus, a scheme that tricks people into pasting commands into their own computer behind fake 'I am not a robot' checks.[26] Version 4.0.9.8 fixes the flaw.[26] Why it matters — Because the Siemens file is genuine, Siemens tells defenders to look for odd behaviour and folders, not for a known bad file.
[26] It says none of its own systems or update channels were broken into.[26] -
17
Urgent fixes for a web server
F5, a US company that sells web and network software to more than 23,000 customers, released urgent fixes outside its normal schedule.
[27] They close several flaws in NGINX, its web server software, three of them critical, that could crash a server or, in some setups, let an attacker run code.[27] Only servers using settings that are not the default are at risk.[27] F5 has not reported any attacks.[27] Why it matters — CISA, the US cyber-defence agency, has listed eight F5 flaws as used in real attacks since 2021, four of them by ransomware gangs.
[27] -
18
Microsoft update breaks company email
Microsoft says its October security update for Windows Server can stop some company email servers running Exchange from working.
[28] Staff may be unable to sign in to Outlook on the web, and emails can stay stuck in the outbox.[28] The update set a new limit on a part of Windows that busy mail servers lean on, and Microsoft has published a manual workaround.[28] Separately, it fixed game and app crashes caused by a September update.[29] Why it matters — Microsoft says the fault is in Windows, and tells customers not to remove their Exchange security updates.
[28] Busy servers are hit hardest, and some may never see a problem.[28] -
19
Drone strike disrupts Russia's biggest social network
A Ukrainian drone strike on the night of 18 October hit a data centre near St Petersburg run by VK, which runs Russia's biggest social network.
[25] On Sunday, Downdetector, a site that collects outage reports, logged about 6,600 complaints about VK.[25] NetBlocks, which tracks internet connections, said VK's traffic fell to 56% of normal.[25] VK says its services kept working and has not said how much was damaged.[25] Why it matters — Ukraine has struck Russian data centres several times this month, and Russia has struck Ukrainian ones in recent weeks.
[25] Datagroup, a Ukrainian telecoms firm, said on Monday that a Russian missile destroyed equipment at one of its sites in Kyiv.[25] -
20
Failed fixes jump tenfold
Terra Security, a company that tests systems with AI, checked public records of software fixes back to 2019.
[30] Until 2025, about 0.2% of fixes listed on GitHub, a site where code is shared, needed a second warning within 90 days because the fix did not work.[30] In 2026 that rose to 2.3%.[30] The Register links the jump to the boom this year in AI tools that find flaws.[30] 2019-20250.2%20262.3%Share of fixes that failed and needed a follow-up warning within 90 days. Why it matters — Flaws whose fixes were later got round were about four times as likely to end up on CISA's list of flaws used in real attacks.
[30] Many failed fixes worked by banning certain inputs, and an attacker only needs one the list missed.[30] -
21
Attacked flaws take three months to fix
Moody's, a firm that rates how likely companies are to repay their debts, says flaws already used in attacks take a median of 87 days to fix.
[31] So half are still open three months after they are made public.[31] Flaws with no known attacks take 107 days.[31] The report uses February 2026 data from Bitsight, a company that measures firms' security from outside.[31] Flaws used in attacks87 daysFlaws with no known attacks107 daysHalf of each kind is fixed faster than this, and half slower. Why it matters — Moody's says firms that fix slowly carry the most risk for lenders.
[31] It adds that cheaper AI attacks could make small and mid-sized firms worth targeting.[31] -
22
Senator asks for hidden website names
Ron Wyden, a US senator, asked CISA, the US cyber-defence agency, to switch on a protection called Encrypted Client Hello across government websites.
[35] Ordinary web encryption does not always hide the name of the site a person visits.[35] His office found at least 350 government sites without the protection, including an FBI tip line and a suicide-prevention line for veterans.[35] Why it matters — Without it, someone watching an internet connection could see who visits a whistleblower or crisis page, Wyden wrote.
[35] CISA did not immediately reply.[35] -
23
Faster police access to website owners
ICANN, the body that coordinates the internet's web addresses, is testing a way for police to ask urgently who owns a website.
[36] Privacy laws now often keep owners' details out of WHOIS, the public register of who holds each address.[36] Requests would be allowed only for threats to life, to vital services, or of child abuse.[36] It is testing the system with the FBI and Interpol.[36] Why it matters — The company that sold the address would have 24 hours to answer, or 72 for hard cases, and could refuse.
[36] ICANN wants each police force to name the few staff allowed to ask.[36] -
24
New list of attacker methods adds AI tricks
MITRE, a US non-profit whose ATT&CK list names the methods attackers use, released version 20 last week.
[32] It adds entries on tricking AI programs with hidden instructions, on using a company's own AI helpers to steal data, and on getting round login codes and 'I am not a robot' tests.[32] It also records attacks on control boxes at more than 30 water utilities in Minnesota.[32] Why it matters — Security teams use the list to check which attacks their tools can spot.
[32] Version 20 adds 15 new ways of spotting them.[32] -
25
WhatsApp pages showed chatbot's rules
About 80 WhatsApp help pages carried, hidden in their code, the instructions Meta gives its support chatbot, says Redact, a privacy company.
[37] One told the bot not to suggest contacting a data protection regulator.[37] Others told it not to mention privacy laws such as Europe's GDPR, even when a user raised them.[37] Meta removed the text after Wired asked and called it a bug.[37] Why it matters — The text shows what a company told its chatbot to avoid when people ask about their rights.
[37] WhatsApp says the bot is meant to help with the app, not give legal advice.[37] -
26
Map shows number-plate cameras
MapQuest, the online map service, has added a layer showing known cameras that read car number plates.
[39] It covers cameras from Flock Safety, Motorola, Axon and others, and shows which way each one faces.[39] MapQuest has not said where the locations come from; ZDNet's reviewer thinks users gathered them.[39] Why it matters — The reviewer counted more than 30 such cameras in a town of 80,000 people.
[39] Flock is fighting maps of its cameras made by members of the public.[39] -
27
Anthropic buys a data-privacy start-up
Anthropic, the company behind the Claude AI, has bought Opaque Systems for an undisclosed sum.
[38] Opaque built ways to use private data in AI while the computer's own chips keep it sealed off.[38] Its team is joining Anthropic, and Opaque is closing down.[38] Why it matters — Opaque has not said what happens to its customers, which included ServiceNow and Accenture.
[38]
Proving who hacked you means showing how you caught them
WhatsApp would not tell a US judge how it spots NSO's spyware accounts, so the judge said it had not proved NSO broke the court's order.
The twist
WhatsApp's way of spotting NSO works only while NSO does not know it. Showing it to the judge would have taught NSO what to change.
The picture
How it works
- A company spots an attacker by signs only it knows to look for
- To punish the attacker, a court asks to see those signs in detail
- Signs shown in public tell every attacker what to hide next time
- So the company holds the detail back
- The case gets weaker, and the attacker keeps working
The same force, elsewhere today
Where this chain is also running, in today's other stories.
-
The grocery giant confirms it was attacked
Ahold Delhaize says it will not share details that could weaken its defences, so shoppers learn little about what happened.
-
Spyware found on a Serbian rector's phone
The Serbian group that found the infection would not say when it found it, citing a sensitive investigation.
-
One hacker, AI programs, 100 websites
NorForge could watch the hackers only through their own open AI tool, and publishing the report and taking those tools down ends that way of watching them.
-
Siemens tool used to slip in malware
Siemens chose to publish the signs of infection, which helps defenders and also shows attackers what is being watched.
Where you've seen this
Police informants
naming a secret source in court can end that source's use in every other case
Virus scanners
publishing exactly what a scanner looks for lets malware writers change just that
Exam boards
explaining how cheating is caught teaches cheaters what to avoid
The catch
This holds only while the method is hard to replace. A company with several ways to spot an attacker can give one up in court and keep the rest.
And the whole of it
WhatsApp's users, Serbia's critics and Ahold's shoppers all rely on defenders who decide, case by case, whether to keep their tools or show their proof. Most of the people affected never see that choice being made.
What is really going on
Most of today's break-ins happened at a company that holds other people's data or keys: ETS held California's pupil records, ConnectWise held a key to its customers' email, OpenCode stored its users' AI keys, and FirstVDS ran its customers' servers.
Why it works on us — A line such as 'no evidence that data was taken' sounds like an all-clear, but Ahold Delhaize, Nikkei and ConnectWise each said it while their checks were still going on.
Who gains
-
NSO Group, maker of Pegasus spyware
— It faces no contempt penalty, because WhatsApp would not show the court how it linked 23 accounts to NSO.
[16] -
US AI companies selling to government
— With DeepSeek banned from civilian agencies, those agencies cannot use a cheaper Chinese model, The Register notes.
[20] -
Hackers using AI programs
— One hacker got into more than 100 sites for about $50 of AI use each, with fewer than 100 characters of instructions.
[14] -
Conduent
— If the court approves the $80 million deal, the company can no longer be sued over the 2024 hack, and it says cyber insurance means the deal will not hurt its finances.
[33]
Who pays
-
California pupils, including those with disabilities
— Their names, addresses and, for some, disability and test details were copied from a company they never chose.
[1] -
OpenCode users
— Some found more than $1,700 of AI charges run up overnight on their stolen keys.
[6] -
Albert Heijn's online shoppers
— Home deliveries stay stopped until at least Thursday, at least 11 days in all.
[5] -
Recruitment agencies on Bullhorn
— Most of their job, onboarding and payroll tools are down with no date to return.
[8] -
Plex owners who had not updated
— Their home servers were made to mine cryptocurrency and carry strangers' internet traffic.
[13] -
Critics of Serbia's government
— Phones of a university head and a corruption prosecutor were targeted with spyware days before the 25 October vote.
[17] [18]
What nobody knows yet
Open questions from across today’s stories — ours included.
-
01
How many California pupils' records were copied.
The state has given no count; it says the files cover every pupil of the last three school years, and the schools had more than 5.7 million pupils in 2025-26.
[1] -
02
How the ETS login was stolen, and by whom.
Neither ETS nor the California Department of Education has said, and nobody has claimed the attack.
[1] -
03
What ETS did to make it less likely the data is published.
ETS says it took steps with police and security experts to reduce that risk, but has not said what they were.
[1] -
04
What kind of attack hit Ahold Delhaize, and whether a ransom was asked.
The company says it will not share details that could weaken its defences.
[3] -
05
Whether any of Bullhorn's customer data was taken.
Bullhorn says its investigation has not yet found out.
[8] -
06
How many companies actually lost email or files through the stolen ConnectWise key.
Huntress counts more than 300 sign-ins that worked, but a sign-in did not always mean access, and ConnectWise says it cannot see inside its customers' systems.
[23] [24] -
07
What the AI-driven hacker took from the 100-plus websites.
Reuters could not find out how much data left the sites, and the researcher's report does not name the hacker.
[14] -
08
When the US ban on DeepSeek actually began.
CyberScoop reports an order letter dated 26 June; the BBC says the ban took effect at the end of July.
[19] [21] -
09
Which consulting firm's executive the FBI suspects of running the extortion site.
KrebsOnSecurity has not named the firm because nobody there has been charged.
[34] -
10
How many FirstVDS customers lost their files.
FirstVDS says only part of its systems was hit, but has given no number.
[22]
Plex had shipped fixes for the four flaws in its home media server by 1 September, six weeks before working attack code was published, so servers that had updated were not open to it.
Also true today
- Conduent agreed to pay $80 million to settle the lawsuit over its 2024 hack, and people with documented fraud losses can claim up to $3,500 each.
- Albert Heijn's website and app are back online, and the Belgian supermarket chain Delhaize has restarted home deliveries.
More from Cybersecurity
Across the beats