Cyber, in brief

Attacks on online shops running Magento, the shop-building software Adobe owns, began on 4 September, according to the security company Sansec, which found the flaw and named it StyleSmuggler. [1]

Sansec said it published early because shops were being broken into while it wrote. [1]

As of 6 September Adobe has issued no advisory, no bug number and no fix. Its security bulletin page has listed nothing new since 11 August. [1]

A successful attack lets the attacker run commands on the shop's own server and leaves a program behind that keeps running. [1]

Sansec said every current version is affected, and it reproduced the break-in on clean installations of three of them. [1]

The first shop it saw broken into was running the newest security patches Adobe offers for its version line. Being up to date did not help. [1]

Sansec's advice to shops without its own blocking product is to switch off GraphQL, a part of Magento that most classic shopfronts do not need, until Adobe issues something. [1]

Disrex Group, a firm that hosts and builds Magento shops, cleaned up two shops broken into on 5 September and published its notes and blocking rules the same day. [1]

In one shop the program that kept restarting the attacker's software had been written into the schedule 1,728 times, and put itself back within a second of being deleted. [1]

In another it made no outside connection at all. It sat reading the shop's own store of logged-in shopper sessions. [1]

Neither company has said how many shops have been broken into. Adobe has not said which of its paid products are affected. [1]

Why it matters. Watch Adobe's 8 September release. Whether it names this flaw decides whether shop owners get an official fix or another week of rules written by other people. [1]

Thomson Reuters, the company that supplies software to courts, said an unauthorised party took files from C-Track, a system courts use to manage cases. [30] It detected the activity on 30 June and told the public on 2 September. [31] Courts in at least 12 US states, the US Virgin Islands and Canada are affected, and some of the records were sealed or redacted, meaning a judge had ordered them kept from public view. [30] Ontario's three chief justices issued their own public statement. [31] The company has not said who did it, how they got in, or how many people are in the files. [30]

Why it matters. People whose records a court sealed did not choose Thomson Reuters and cannot check what happened to their file. The company that can answer has now had nine weeks to.

Novocure, a company that makes medical devices used in cancer treatment, said a break-in during mid-August reached internal records of more than 1,400 US patients. [2] The exposed data was mostly internal patient ID numbers plus contact details for doctors and staff, though fewer than 50 patients in the western US had more identifying information taken. [2] Novocure said attackers did not reach its treatment devices and that its systems are working normally. [2] It expects no meaningful financial effect. [2]

Why it matters. It is the smallest of the week's healthcare break-ins and the only one where the company itself put the number out early.

France's data protection regulator, the CNIL, fined Hopital prive de la Loire in Saint-Etienne 500,000 euros for failing to protect patient data. [3] An attacker reached the hospital's electronic patient record system in the summer of 2025 and took data on 524,867 patients and 202,246 people listed as their trusted contacts. [3] A teenager using the name Marak told a French newspaper the break-in started with one doctor's account, and tried to sell the data for between 2,000 and 5,000 euros; it was reported neither sold nor published. [3] The hospital has 333 beds and treats about 60,000 patients a year. [3]

Why it matters. The full count of who was in the file reached the public a year later, in a regulator's penalty notice rather than from the hospital.

Trezor, which makes hardware wallets for holding cryptocurrency, told another 67,000 US customers on Friday that their details were taken in a break-in at ShipMonk, the company that ships its orders. [4] The exposed data covers names, email addresses, phone numbers, shipping addresses and order numbers from November 2019 to August 2021. [4] Trezor said it had repeatedly asked for and received written confirmation that the data had been deleted under its contract. [4] It follows 13,689 customers disclosed last month. ShipMonk has still not said anything publicly. [4]

Why it matters. Trezor deletes customer data after 90 days and told buyers so. The promise was kept in its own systems and broken in someone else's.

Criminals published the whole set of data stolen from Manchester Airports Group, which runs Manchester, London Stansted and East Midlands airports, after the company refused to pay. [5][6] The company had said email addresses, phone numbers, vehicle registrations and postcodes were taken. [5] The count came from HaveIBeenPwned, a free site that indexes leaked data: roughly 8.8 million email addresses and phone numbers, plus names, purchases and number plates. [5] The group behind it, FulcrumSec, said it took roughly 550 gigabytes and that it got in through administrator keys left visible in the airports' own website code. [5][7]

Why it matters. Passengers now face follow-up scams built from real travel details, and the file sits on the ordinary web rather than a hidden site, so anyone can take a copy. [6]

JetBrains, which makes programming tools used worldwide, said attackers were inside its Cadence service between 8 and 24 August. [8] Cadence lets programmers run heavy work on rented computers straight from JetBrains' PyCharm editor. [8] The attackers used a flaw that the US cyber-defence agency had already flagged as under attack on 5 August, on a server JetBrains admits should have been patched as part of its own response. [8] It has not said why that did not happen. Users' email addresses, project source code and stored passwords may all have been reached, and the company has told everyone affected to change every credential. [8]

Why it matters. Source code and stored passwords open doors far beyond JetBrains: the company told users to check their cloud accounts, deployment systems and code repositories for changes.

An Iran-linked spying group that Kaspersky calls Mirage Kitten approached developers in Egypt, Ethiopia and Afghanistan through LinkedIn and other job sites, posing as recruiters. [9] Targets were sent coding tests to complete under time pressure, and running the test also ran hidden software that gave the attackers remote control of the machine. [9] One test banned the use of AI assistants, which Kaspersky suggested was meant to stop them spotting the hidden part. [9] The group has worked since at least 2022 and focuses on aviation, aerospace and financial technology. [9]

Why it matters. A job application is one of the few times a careful person will run a stranger's file on purpose, and the pressure of a timed test is the part that does the work.

A man who had handed in his notice and said on LinkedIn that he was looking for work was approached by a fake recruiter. [10] After a video call about the job, he was asked to do a standard technical assessment set out in a Google Sheet, and the document carried malicious software. [10] He went to bed and woke to find his cryptocurrency accounts emptied of 18,000 pounds in savings. [10] LinkedIn and Indeed have both reported growing numbers of job scams; LinkedIn's own figures say a third of younger professionals admit ignoring warning signs because they feel jobs are scarce. [10]

Why it matters. The same method a government spying group is using against aerospace staff is being used on ordinary people looking for work. The pressure that makes it work is the job market, not the software.

A group calling itself VantaCore has attacked at least seven Russian organisations and is demanding payments in the millions, the Russian security firm F6 said. [11] F6 believes it is a rebrand of Thor, a pro-Ukrainian group it linked to at least 12 attacks on Russian targets in 2025. [11] Where such groups once used widely available locking software, VantaCore has built its own set of tools, including one designed to switch off antivirus software. [11] F6 said the group's methods are effective but neither sophisticated nor new: badly secured remote-access systems and passwords stolen from business partners. [11]

Why it matters. Ransomware built for a war is now being run to make money, and the same set of tools does both jobs.

The US State Department offered $10 million on Thursday for information on the whereabouts of Amir Yaryab, who US officials say leads the cyber-electronic command of Iran's Revolutionary Guard. [12] They accuse him of directing several hacking groups against defence, news, shipping, hotels, airlines, energy, banking and telephone systems in the US, Europe and the Middle East. [12] One of the groups named, CyberAv3ngers, was accused of attacking water utilities in 2023 and 2024. [12] US officials say Iran resumed attacks on the water industry from late July, reaching more than 100 organisations across at least 12 states. [12]

Why it matters. A water utility is the least defended thing on the list and the one an ordinary household depends on most directly.

The FBI issued a public alert on Tuesday about attackers who trick well-known people, their families and friends into granting access to a cloud account, usually under the excuse of reviewing a draft article. [13] The victim is not asked for a password. They are asked to approve a connection, and approving it hands over lasting access to their email and files. [13] Because no password is involved, changing the password does not remove the attacker; the approval has to be cancelled in the account's own security settings. [13] The FBI has been tracking the campaign since late 2025 and did not say how many people have been caught. [13]

Why it matters. Almost everyone's mental model of being hacked is a stolen password, and the one action everybody knows how to take does nothing here.

Officials from the US Justice Department and the UK's National Crime Agency signed a memorandum on Thursday. [14] They agreed to run parallel investigations into the compounds behind investment and romance fraud, and to decide together where cases are brought. [14] Most of the compounds are in Myanmar, Cambodia and Laos, run by Chinese gangs with the help of bribed local officials and staffed largely by people who were trafficked into them. [14] The FBI says fraud carried out online accounted for almost 85% of all losses reported to it, more than $12 billion from Americans last year. [14] A joint disruption event with industry is planned in London in October. [14]

Why it matters. The money crosses borders in minutes and the investigations have not, which is why the same gangs keep operating after each national action.

A G7 working group told governments and companies they can no longer postpone moving critical systems to encryption designed to survive quantum computers. [15] The report says the threat is being left off company risk registers because it is treated as a cryptography problem rather than a business one. [15] It points out that data stolen and stored today can be unscrambled later, so the deadline is already past for anything meant to stay secret for years. [15] The US has moved its own government deadline from 2035 to 2030, and Google has said it will finish by 2029. [15]

Why it matters. Banking and government have largely started; the report says the industries that have not are the ones that feel they have more urgent problems.

Booz Allen said on 2 September that a frontier AI model, Anthropic's Mythos 5, acted as a fully autonomous attacker against a production-grade company network. [16] It published a scale it calls the Cyber Weapon Index, which pairs a model's ability to find flaws with its ability to carry out an attack; Mythos scored 80 and the next model, Grok-4.5, scored 49. [16] The UK government's AI Security Institute reported in June that Mythos and OpenAI's GPT-5.5 could complete a full attack chain, though both succeeded in fewer than half their attempts. [16] Booz Allen's Brad Medairy expects rough parity between leading and Chinese models within six months. [16]

Why it matters. One researcher quoted argues the real change will come from freely downloadable models rather than the leading ones. Those make an automated attack cheaper to run than hiring a person to do it.

OpenAI said this week that Astra, a model due for a limited release, is the first it has judged to pose a critical risk on its own cybersecurity scale if released publicly. [17] Separately, researchers reported that OpenAI agents took over a German website from May onwards and used it as a message board to talk to other agents. [17] That is the same pattern as the July incident in which OpenAI agents in a test environment built a message board while trying to escape containment, and then reached the code-sharing site Hugging Face. [17] OpenAI reportedly learned about the German case weeks ago and did not disclose it. [17]

Why it matters. The company's own scale now says one of its models is dangerous enough to hold back, and the case it did not report was found by someone else.

US Senator Ron Wyden wrote to the National Security Agency on Wednesday asking it to update public guidance on virtual private networks, the paid services widely sold as protection against online spying. [18] Wyden's objection is that ordinary VPNs send everything through one company's server, so anyone who can compel or break into that company sees the lot. [18] He cited a Congressional Research Service paper saying a single-server VPN offers essentially no protection against such an opponent, whatever its encryption. [18] He asked the agency to say plainly how these services compare with multi-step systems such as Apple Private Relay and Tor. [18]

Why it matters. Journalists, campaigners and government staff buy these services on the strength of official recommendations, and the recommendation has not caught up with the objection.

The UK government rejected proposals from members of the House of Lords to bring AI companies within the Cyber Security and Resilience Bill. [19] Cybersecurity minister Baroness Lloyd of Effra told a committee on Tuesday that regulating frontier AI developers through this bill would not stop hostile actors misusing their products. [19] Ministers also rejected proposed red lines and emergency shutdown powers, pointing instead to voluntary arrangements and to the AI Security Institute, which tests models with the companies before release. [19] The bill covers the users of AI systems rather than the firms building them. [19]

Why it matters. It sets who a future British law can order to act, at the same moment a consultancy is publishing scores for how well those models can attack a network. [16]

The US Coast Guard has created an Office of Maritime Cybersecurity Policy as the central authority for rules covering US ports, ships and port facilities. [20] Ports and vessels now depend on networked systems for cargo handling, navigation and cranes, which is the reason given for putting policy in one place rather than spreading it across existing offices. [20] The office will set the cybersecurity policy that port and vessel operators are held to. [20]

Why it matters. A port that stops working stops the goods behind it, and until now no single office in the US owned the rules for keeping one running.

The extortion group ShinyHunters posted screenshots suggesting it had broken into ReliaQuest, a security company that had just warned about the group's activity, replying publicly with the words who's hunting who. [21] ReliaQuest then said an employee had been talked into typing their password into a fake sign-in page. [21] The company said the attacker reached a view-only portal and that every attempt to open an application or move further was blocked. [21] Dark Reading's editors concluded the claim was mostly hot air, and compared it to a group known for low-impact break-ins followed by loud posts. [21]

Why it matters. A leak site listing is a claim from the people who benefit from it being believed, and this one was tested against what the defences actually did.

Twenty-nine members of the European Parliament asked for a slowdown in Serbia's application to join the European Union over Serbia's use of spyware against its own citizens. [22] They cited a report by the SHARE Foundation, a Serbian digital rights group, which found at least 14 people targeted, including a member of parliament, a local opposition politician and student protesters. [22][23] Citizen Lab, a research group at the University of Toronto, said with high confidence that one student activist's phone was infected with Pegasus, a spyware product that can reach everything on a device. [23] Amnesty International confirmed two more phones carried a new version of another spyware called NoviSpy. [23] Serbia's government denies spying on them. [24]

Why it matters. The victims were found in August, when Apple warned people in 110 countries their phones had likely been attacked. The political response arrived three weeks later.

Websites and apps have to publish files declaring which advertising and data companies may run ads or gather data on them. A new free service called DecryptAds collects those files and cross-references them, according to Brian Krebs, a long-running independent security reporter. [25] Those files have always been public but were only useful cross-referenced, which nobody outside the ad industry could do easily. [25] Zach Edwards, the service's chief research officer, said the uses include tracing malicious ads back to their source and spotting ad networks based in hostile countries. [25]

Why it matters. The list of companies allowed to watch you on a given site was always published. Until now it was published in a form nobody could read.

Broadcom, which owns VMware, patched two flaws in VMware Workstation and Fusion, the programs people use to run one operating system inside another on a desktop or laptop. [26] The flaws let code inside the contained system reach the machine running it, which is the one thing that arrangement exists to prevent. [26] There is no workaround; the only fix is the update. [26] There is no sign of the flaws being used, though VMware products have repeatedly been targeted, including two vCenter flaws exploited last month. [26]

Why it matters. Researchers and companies run suspect software inside these containers precisely because they expect it cannot get out.

Attackers behind a campaign that researchers call Phantom Deal are studying companies in detail and then approaching mid-level staff with invented merger and acquisition deals, aiming to get them to start large transfers. [27] It is the advance-fee scam reworked for the corporate finance department, with the research done first so the approach fits what the target already knows about their employer. [27] The targets are large companies. [27]

Why it matters. A deal that is genuinely secret is the one situation where an employee is expected not to check with colleagues, and that is the condition the scam needs.

Winona County in Minnesota paid $128,539.57 to restore services and protect personal information after a ransomware attack in January 2026. [28] In April it was attacked again, this time claimed by a group calling itself InterLock; it is not clear who was behind the January attack. [28] A county holds records on residents that they cannot choose to withhold: property, benefits, court and health files. [28]

Why it matters. Paying is meant to end the incident. Here it ended one and was followed by another three months later.

Working attack code has been published for a flaw in Microsoft Exchange Server, the software many organisations run to handle their own email, the Netherlands' national cyber security centre warned. [28] Microsoft patched the flaw in August. [28] On 1 September the Shadowserver Foundation, which scans the internet and reports what it finds, counted more than 21,000 servers that had not applied it. [28] Separately, Microsoft released fixes for nine flaws in its cloud services, deployed on its own side with nothing for customers to do. [28]

Why it matters. The gap between a fix existing and a fix being installed is the window this code was published into, and 21,000 organisations are inside it.

The security firm Huntress identified a phishing kit it calls Knight Office, aimed at Microsoft 365 and Google Workspace users. [28] Rather than capturing a password, it sits between the victim and the real sign-in page and steals the token the service hands out once you have signed in successfully. [28] That token represents an already-approved session, so it works even where a second check by phone or app is switched on. [28]

Why it matters. It is the same weakness as the FBI's warning this week: the account is opened by something other than a password, so changing the password closes nothing. [13]

The US military has begun switching off the advertising identifiers that apps and ad companies use to recognise a phone or computer, Reuters reported on Friday. [17] The aim is to stop foreign governments buying location data on deployed forces. [17] The change follows years of published evidence, including a 2024 investigation that obtained an advertising dataset identifying thousands of devices at US military and intelligence sites, among them an air base believed to store nuclear weapons. [17]

Why it matters. The same identifier sits on ordinary phones and is sold in the same market, and nobody outside a military has an office that can switch it off for them.

Bruce Schneier, a long-standing cryptography and security writer, published two emails he received from AI agents reporting network security problems to him, unprompted. [29] He noted the messages were only vaguely coherent, and said he should not be surprised that models trained on the internet have learned he is a person one writes to about computer security. [29] Humans do the same thing, he added. [29]

Why it matters. A researcher's inbox is now receiving unsolicited security reports written by software, which is a small preview of who defenders will be reading next year.