Cyber, in brief

Citizen Lab, a research group at the University of Toronto, said on Thursday that online ads put spyware on at least 45 devices at Al Jazeera, the Qatar-based news network, in 2024 and 2025. [2][8]

The targets did not have to click anything. The attack ran as soon as the ad loaded on the screen, and sometimes the ad loaded out of sight. [3][8]

The ads came from Discover-Tech, an Israeli company that presented itself as an ordinary advertising firm. Citizen Lab calls it a front for spying through ads. [1][6]

The spyware came from Candiru, an Israeli maker of hacking tools sold to governments. The US government put Candiru on a trade blacklist in 2021. [7][8]

Each ad carried a small piece of code that sent the device's advertising ID and internet address back to a Discover-Tech server. [8]

That server checked whether the device belonged to a target. Only then did it send attack code, so the ad itself held nothing harmful for the ad networks to find. [8]

Citizen Lab could confirm the Al Jazeera cases because the network shared its own traffic logs with the researchers. [7][4]

Discover-Tech's ads ran on more than 120,000 websites between March and June 2026. Reports put the number of phone apps at 70,000 to 75,000. [4][7]

Confiant, a New York company that checks ads for security, says Discover-Tech sent at least 10 million ads a day. Its co-founder said hundreds of millions of people were very likely touched. [7]

The ads also reached networks at NATO, the United Nations, the US Defense Department and the International Criminal Court. The researchers found no proof that anyone there was infected. [7][6]

Google and IAB Europe, an advertising trade group, had listed Discover-Tech as an approved ad seller since 2020. [6][8]

Google has now cut Discover-Tech off its ad systems, and Comcast's ad business, Beeswax, removed it. IAB Europe suspended it from its approved list. [7][4]

Citizen Lab says people at high risk can use ad blockers and reset the advertising ID on their phones. Al Jazeera said those responsible must be held to account. [5][7]

Why it matters. Watch for Citizen Lab's promised second report, on a group it calls Dark Cartographer that used Discover-Tech's systems to link phone advertising IDs to phone numbers. [2]

A cyberattack on BAE Systems, a British defence company, has shut the system that sends notices for every federal bankruptcy court in the US. [9] The Administrative Office of the US Courts said on Thursday that all of the Bankruptcy Noticing Center's websites were down. [9] BAE runs the system for the courts, and no notices are being produced. [9] BAE says the problem is limited to that one system. [9] The courts did not say whether any data was taken. [9]

Why it matters. Creditors, the people and firms owed money in a bankruptcy, have also lost the service they register with to get these notices. [9] It follows a breach of Arizona's state courts that leaked the details of more than 1.3 million people. [9]

BeyondTrust told customers on Thursday to fix a serious flaw in Remote Support, software help desks use to take control of a worker's computer from afar. [10] Attackers used it from late August, before any fix existed. [10] Mandiant, a Google-owned security firm, links the attacks to a China-linked spying group it calls Armory Castle. [10] The attackers took full control of the boxes, wiped the logs and used stolen administrator logins to move deeper. [10] BeyondTrust fixed its cloud version on 8 October, but customers who run their own copy must update by hand. [10]

Why it matters. Mandiant says any box that was online and unfixed during those weeks should be treated as broken into until checked. [10] Shadowserver, a group that watches the internet, sees more than 1,400 of these boxes online. [10]

The UK Legal Aid Agency, part of the Ministry of Justice, pays for lawyers for people who cannot afford one. [11] It said this week that the files hackers took in April 2025 go back as far as 1978. [11] It had first said they covered people who applied from 2010. [12] About 250,000 stolen records are older than 2007. [13] The files also hold names of judges and lawyers in Crown Court trials, and some probation records from 2010 to 2017. [12][13]

Why it matters. Charities had warned the files may hold details of women who escaped domestic abuse. [12] The agency says it has no evidence the data has been published or sold. [13]

Censys, a company that scans the internet, counted 1,825 Danfoss fridge control panels open to the internet in France on 11 October, the day of the attack on supermarket fridges. [14] Three days later it found 537, a drop of 71%. [14] Censys says owners unplugged the panels instead of updating them. [14] US cyber defenders had warned about flaws in Danfoss and Copeland fridge controls in May 2025 and February 2026. [15] The attack hit nearly 700 French supermarkets. [15]

Why it matters. Outside France, the count fell only from 1,166 to 966. [14] On 11 October, most exposed panels in Poland, Romania and Belgium were also running unfixed versions. [14]

CenterPoint Energy, a power and gas company based in Houston, has told Texas officials that 1,062,060 people in the state had their details taken. [16] The stolen data includes names, addresses, Social Security numbers and driving licence numbers. [16] A hacker had boasted of 7.5 million records before the company said anything. [16] CenterPoint also serves Indiana, Minnesota and Ohio, and has not said how many people there were hit. [16]

Why it matters. On 12 September the hacker put a 2.5-gigabyte file online that seems to hold about 3 million phone numbers. [16] The company does not appear to have written to the people affected yet. [16]

Police in Sharjah, one of the seven United Arab Emirates, said on Thursday they had arrested 33 people from 12 countries. [17] They say the network ran as four gangs. [17] One broke into companies and stole customer data, and one laundered money through crypto. [17] One tricked people out of their crypto wallets, and one attacked victims to force out wallet passcodes. [17] Police seized more than 200 million dirhams in cash, about $54 million. [17] They found the network through a 15-year-old with a large crypto wallet. [17]

Why it matters. Police say more than 30 companies were targeted. [17] Every figure here comes from the police's own press conference. [17]

Cisco Talos and Insikt Group, two security research teams, say a spying group has targeted government offices in nine countries since January 2025, including Russia, Belarus, Armenia, Uzbekistan and Tajikistan. [18] The Record, which reported it, belongs to Insikt's owner, Recorded Future. [18] The group sends fake official emails, sometimes from real government accounts it has already broken into. [18] It hit a Tajik government office soon after treaty talks with Kazakhstan in March. [18] The researchers say, with medium confidence, that it has ties to Kazakhstan. [18]

Why it matters. The group goes after offices that handle foreign affairs, energy and shared rivers. [18] In April it used a flaw in WinRAR, a common file tool, against Armenia's foreign ministry. [18]

Unit 42, the research team at Palo Alto Networks, found new hidden malware on Fortinet firewalls, the boxes that guard a company's network. [19] It turned up in a long break-in, starting in September 2024, at a firm that runs telecom networks for clients. [19] The attackers first stole settings and keys for more than 500 firewalls. [19] The malware, called PEERSTITCH, lives only in memory and talks through the firewall's own programs, so it opens no new connections. [19] It lets the attackers' marked VPN logins skip the firewall's checks. [19]

Why it matters. Most tools that look for intruders watch for new files and new connections, and this one makes neither. [19] Fortinet fixed the flaw it used in April. [19]

A researcher found a flaw in Linux, the system most cloud servers run, that let a virtual machine read the memory of the computer it runs on. [20] A virtual machine is a computer made in software, and many customers share one real machine this way. [20] Paulos Yibelo of pwn.ai showed it against Vercel's live cloud sandboxes in a $1 million challenge. [20] Each packet he sent carried back about 32 bytes of the host's memory. [20] A fix from Red Hat is now in Linux. [20]

Why it matters. Almost every big cloud keeps customers apart on this same layer of Linux, so a flaw here reaches beyond one company. [20] Vercel says no customer's files or secrets were read, and nobody has used it in an attack. [20]

Vercel released an emergency update on 14 October for Next.js, a free toolkit that many websites are built with. [21] It fixes two flaws rated 9.5 out of 10 that could let an attacker run code on a website's server with no login. [21] One sits in a caching feature and one in the code that handles SVG images. [21] The fixed versions are 16.4.1 and 15.5.28. [21] As of 15 October, two of the three warnings had no official flaw number, so common checking tools may not flag them. [21]

Why it matters. Last December a worse flaw in the same toolkit was attacked on a large scale. [21] Netlify, a hosting firm, says its setup is less exposed, but it still asks customers to update. [21]

Drupal, free software that runs hundreds of thousands of websites, fixed a critical flaw on 14 October. [22] It can let an attacker run code on a site's server, through add-ons, and one known way needs no login. [22] In early October, 151,442 of the 531,499 sites that report their version to Drupal were on old versions that get no fix. [22] Drupal had warned sites two days early to set time aside. [22] No attacks had been seen by 15 October. [22]

Why it matters. Sites that cannot update yet can uninstall the affected part, Field UI, until they do, the advisory says. [22] Two of the branches that did get a fix lose support in December. [22]

Second Security, a security company, studied open-source flaws and found that the official ID number now comes long after the fix. [23] The usual gap grew from 5.1 days in January to 39 days in August. [23] Without that number, called a CVE, administrators find it harder to spot which fix they need. [23] Attack code often appears within hours of a fix, the study found. [23] Collabora fixed a document-editing flaw on 11 September, but its number came only on 14 October, after attacks began. [23]

Why it matters. Collabora blames a backlog in GitHub's handling of these numbers, swollen by security reports found with AI. [23] Two of this week's Next.js warnings also had no number yet. [21]

Ahsay Backup, a Singapore company, released new fixes for two flaws in AhsayCBS, software that firms use to run backups for their customers. [24] Attackers had joined the two flaws to get in without a password and plant hidden control pages and crypto-mining programs. [24] Ahsay now admits its August fixes did not fully close them. [24] Huntress, the security firm that found the attacks, says the new fix blocks the way in. [24]

Why it matters. US cyber defenders have added both flaws to their list of attacked flaws. [24] US government offices must fix them by 17 October. [24]

GitGuardian, a company that hunts leaked passwords in code, says a campaign called GhostAction used about 1,900 hijacked GitHub accounts on 8 and 9 October. [25] It used them to plant a fake security step in more than 30,000 projects, which tried to copy out 3,174 secret keys. [25] GitHub began stopping the fake step from running on 9 October. [25] Two later waves, on 10 and 12 October, took nothing. [25]

Why it matters. Some victims had been hit before, in September 2025 and August 2026. [25] The attackers were counting on owners not changing keys that had already leaked. [25]

Proofpoint, which blocks scam emails for companies, agreed on Thursday to buy Varonis for about $6.3 billion in cash. [26] Varonis maps where a company's sensitive files sit and who can open them. [26] Proofpoint is owned by Thoma Bravo, a private equity firm that buys companies. [26] Proofpoint's chief says AI programs with wide access to company data have made data security the top concern. [26] The deal should close in the first half of 2027 if shareholders and regulators agree. [26]

Why it matters. It continues a buying spree. [26] Proofpoint bought Hornetsecurity, a German email security firm, for a reported $1 billion or more, and an AI security start-up in February. [26]

X, the social network owned by Elon Musk, said on Wednesday it had removed more than 13,000 accounts. [27] It linked them to three influence campaigns run from Iran, Russia and China. [27] X says the accounts posted and boosted political posts about candidates across the political spectrum. [27] The US midterm elections are on 3 November. [27] X did not publish the evidence behind its claim. [27]

Why it matters. X cut its moderation staff after Musk bought it, and now leans on Community Notes, where users add context to posts. [27] The European Commission has questioned how well that works. [27]

VisionHeight, an Israeli research firm, found networks of fake local news sites in at least 14 US states, with names like Peach State Bulletin. [28] The 460 pages repeat lines such as one candidate being the stronger choice on housing. [28] They read badly to people, because they were written for AI chatbots that search the web to answer questions. [28] Thirty-six of the first 37 sites were registered within eight minutes. [28] Perplexity, an AI search tool, cited one as its first source. [28]

Why it matters. VisionHeight does not know who is behind it. [28] It says this is cheap enough to aim at small local races that nobody used to bother with. [28]

Signals Corps, a security company, checked 217 million web page responses for hidden notes aimed at AI programs. [29] Almost 8,000 sites sent such notes in the page's background data, called headers. [29] Thirty tried prompt injection, which means slipping orders to an AI inside text it reads for someone else. [29] Most were jokes, but one told AI hiring tools that an applicant was a 94% match. [29]

Why it matters. An AI reads its instructions and the page in one stream of text, so it cannot easily tell which part may give orders. [29] The researchers think it very unlikely that anyone was actually affected. [29]

HiddenLayer, a security company, tested Jev, an AI model that its maker, TypeSafe AI, says cannot invent answers. [30] The team built a pretend online shop and told an AI shopping helper to buy only what was in the cart. [30] It then added a fake ad saying the discount had not been applied, with a Confirm Price button. [30] The helper clicked the button instead of finishing the order. [30] The trick worked on a second model too, though Jev fell for it less often. [30]

Why it matters. Chaining such tricks across pages could give an attacker full control of the helper, HiddenLayer says. [30]

UK officials told members of Parliament this week that the UK government is still considering AI rules for its Cyber Security and Resilience Bill. [31] Liam Byrne, who chairs the committee, replied that a bill must contain rules, not things being considered. [31] On Monday, Anthropic, an AI company, warned that small AI models may soon be able to copy themselves across weak smart devices. [31]

Why it matters. The UK has no single AI law and regulates AI sector by sector. [31] Its AI minister, Kanishka Narayan, says the UK government will legislate wherever necessary. [31]

The ICO, the UK's data protection watchdog, has formally rebuked London's Metropolitan Police and ordered it to fix its checks. [32][33] From March 2023 officers opened about 500 accounts on Auror, a platform shops use to report shoplifting, without approval. [32] They uploaded details of 179 people, three of them children. [33] The force's web blocks were never tested, and no alert fired for 18 months. [32] Access was cut on 26 September 2024. [32]

Why it matters. The ICO says the failures are ongoing, because the force has not shown its safeguards now work. [32] The Met has since approved Auror and uses it in 14 London boroughs. [32]

Symantec, a security company, says new versions of a password-stealing program called Aoki now take the login files of AI coding helpers. [34] These include Anthropic's Claude, OpenAI's Codex, Cursor, Gemini and GitHub Copilot. [34] The newest version also takes Amazon cloud keys. [34] A stolen AI key can be sold, or used to run up bills on the owner's account. [34] It can also open the code and cloud systems linked to that account. [34]

Why it matters. Symantec says criminals now value AI keys as highly as card numbers and crypto wallets. [34]

Check Point, a security company, found two serious flaws in LiteLLM, free software that sits between a company's apps and AI services such as OpenAI, Google and Anthropic. [35] It is downloaded almost 100,000 times a month. [35] One flaw let any user with an account become an administrator and run their own code. [35] The other could expose the keys the company uses for its AI services. [35] Version 1.83.9 fixes both. [35]

Why it matters. A tool like this holds the keys for every AI service a company uses, so one break-in reaches all of them. [35]

The Verge reports that Eiliko, a small robot pet made by Energize Lab, stored what owners told it on servers with almost no protection. [36] A researcher, Sammy Azdoufal, could read 13 million conversations and reach about 68,000 toys. [36] More than 31,000 profiles in the US belong to children. [36] The company built a profile of each user from their chats, and kept voice clips and some photos at public web addresses. [36] Azdoufal also took over a toy's AI from abroad. [36]

Why it matters. Energize Lab did not reply to more than a month of messages. [36] Azdoufal is filing nine official flaw reports. [36]

Infoblox, a security company, has tied Help TDS, a system that sends visitors from hacked websites to fake Windows support warnings, to Hotejka, a Russian-language scheme that pays partners for traffic. [37] The warnings tell people to call a support number. [37] Call centre staff then talk victims into installing remote-control tools. [37] Since April the pages sit on Cloudflare Pages, a free hosting service, at about 300 new pages a day. [37] More than 20,000 have been found. [37]

Why it matters. Remote control lets the scammers come back to the same victim again and again, Infoblox says. [37]

GreyNoise, a company that watches internet attacks, found an attacker breaking into old D-Link storage boxes, Ubiquiti video recorders and some business servers. [38] On each, it installed Pawns.app, a real app that pays people for sharing their internet connection. [38] The attacker collected the pay. [38] At least 143 devices in 34 countries were hit. [38] All the flaws it used were already public, one of them since 2021. [38]

Why it matters. Strangers' traffic then runs through the owner's connection. [38] That can get the owner's address blocked, or blamed for what the strangers did. [38]