Cyber, in brief

The US Justice Department seized the Telegram channels of Xinbi Guarantee, an online marketplace where scam gangs hired each other. [1]

Telegram is a messaging app. Xinbi ran on it as a set of channels where sellers advertised and buyers paid. [1]

On sale: custom-built fake investment websites, laundering of money taken from victims, and people trafficked into scam compounds to do the talking. [1]

Xinbi's own role was to hold the buyer's payment until the seller finished the job, so that neither side had to trust the other. [1]

Investigators froze $52.8m across 52 wallets in a single day. About $12m of that sat in two wallets Xinbi used to collect payments for its sellers. [1]

The tracing firm Elliptic found the wallets with the US Secret Service. Elliptic puts Xinbi's transactions at $30bn since around 2022. [1]

Every frozen coin was Tether's USDT, a token meant to be worth one dollar. Tether, the company that issues it, can switch off any wallet holding it. [1]

Xinbi answered by moving about $2.8m into USDD, a token with no issuing company and so nobody to ask for a freeze. [1]

The same operation shut 13 scam compounds in Madagascar. Nearly 400 people were arrested there and about 30 alleged Chinese bosses were sent back to China. [1]

Days before, the US Treasury's financial crimes unit FinCEN counted $12.7bn stolen from Americans in cryptocurrency investment scams in a year. [2]

That count came from about 1,300 banks and crypto firms, and the monthly number of reports was rising by nearly 11%. [2]

Adults over 60 were only a quarter of those reports, so the losses are spread across ages. Most victims realised only when asked to pay a fee to withdraw. [2]

Why it matters. Watch whether Xinbi's sellers follow it into USDD, and whether that holds: Elliptic says USDD is partly backed by USDT, so some of the freeze risk travels with it. [1]

Brevo, a Paris email marketing company, was broken into, and the attackers used that access to send mail from its clients' genuine addresses. [4] Customers of Trezor, BitBox and CoinTracking got what looked exactly like each company's own security notice, and several clicked through to near-perfect copies of the login pages. [4] Trezor says the attackers reached its opt-in newsletter list of roughly 347,000 addresses, and that 120 Brevo accounts were hit in all. [6][5] The fake Trezor mail asked people to install an app and type in their wallet backup phrase. [6]

Why it matters. A sender address is the one thing most people use to judge an email, and here it was genuine. [4] For a hardware wallet owner, typing the backup phrase hands over every coin the device holds. [6]

Trezor, which makes hardware wallets for storing cryptocurrency, said on Friday that its August breach reaches 81,000 customers rather than the 14,000 first announced. [7] The extra 67,000 are US buyers who ordered between November 2019 and August 2021. [7] Their names, addresses, phone numbers and order numbers were still sitting at ShipMonk, the shipping firm Trezor used. [7] Trezor says its contract required ShipMonk to delete that data, and that it had written confirmation the deletion had happened. [7] The attackers got in through a flaw in Metabase, a reporting tool ShipMonk ran. [7]

Why it matters. Orders placed seven years ago were exposed only because nobody actually deleted them. [7] The people in that file are wallet owners with home addresses attached, which is the list the phishing above is built from. [7][4]

Check Point, which makes firewalls and VPN gear used by companies, patched two critical flaws in how its products read VPN certificates. [8] A certificate is the file that proves a machine is who it claims to be before a private connection opens. [8] Both flaws can let an attacker who has never logged in run code on the box, though the company says only under conditions it has not described. [8] Check Point found both itself, published them on 9 September and began shipping fixes the same day. [8]

Why it matters. A firewall sits at the edge of a company network, so code running on it starts inside. [8] Check Point assigned the 9.8 scores itself, and a staff member said the certificate flaw could in theory be reached where no VPN is switched on. [8]

Cisco's Talos researchers say two separate groups broke into Secure FMC, the software companies use to manage Cisco firewalls. [9] One group was tied with high confidence to affiliates of the Qilin ransomware gang. [9] They logged in using a fixed password that shipped inside the product, mapped the network with the tool's own features, then encrypted files. [9] The second group used tooling that overlaps with Sandworm, a unit of Russian military intelligence. [9] The US cyber agency CISA ordered federal agencies to patch by 12 September. [10][11]

Why it matters. The same hole served a criminal gang and a state unit within weeks of each other. [9] Cisco's own example log entry is dated 23 July, weeks before the company says it learned of the attacks in August. [10]

Zimperium, a mobile security firm, described Mantax Otax, Android malware spread through installer files hosted outside Google's own store. [12][13] On phones running Android 9 or older it encrypts photos and documents, deletes the originals, and opens a chat window to haggle over the ransom. [12][13] On Android 10 and later a storage rule blocks most of that, so it mainly spies: contacts, call logs, one-time login codes, WhatsApp and Telegram messages, live screen recording. [12][13] A later version adds pop-ups every 600 milliseconds and makes the handset speak the attacker's words aloud. [12]

Why it matters. The harassment is the payment pressure, aimed at people whose only copy of their family photographs is on the phone. [12] A badly configured server left the attackers' own negotiation chats open to researchers. [12][13]

Researchers at Dark Atlas published THost9, an Android remote-access tool that hides its icon and runs as a background service. [14] Its second stage carries a worm aimed at Android Debug Bridge, a developer port that some devices leave open to the internet. [14] The worm expands one address into a range of 65,025 hosts and probes them fifty at a time, then installs itself where it gets in. [14] Where the session is privileged it copies itself into a system folder, which survives an app being removed. [14]

Why it matters. That port exists so a developer can plug in a cable, and devices shipped with it exposed are what this spreads through. [14] A newer build quits the moment it detects Frida, a tool researchers use to inspect running apps. [14]

Google Play's Early Access programme lets people try unfinished apps, and it carries no star ratings and no public reviews. [15][16] Bitdefender, a Romanian security firm, found thousands of deceptive apps using that gap, including fake casino games and reward apps promising PayPal payouts that never arrive. [15][16] One imitation of Grand Theft Auto called Vice Streets: Open World passed a million downloads with no reviews at all. [15] The traffic comes from adverts on TikTok and Facebook, many using AI-made videos of famous athletes and actors. [15][16]

Why it matters. The feature that shields an unfinished app from unfair criticism also removes the first warning a buyer would get. [15] Casino-style apps in the programme also sidestep the licensing rules real gambling apps have to meet. [15][16]

New analysis describes JSCeal, malware spread through fake cryptocurrency trading sites advertised on Facebook and Google. [17] Victims are told to download an installer for TradingView, a charting service, and get the malware instead. [17] It steals session cookies, the small file a browser keeps after you sign in so the site stops asking again. [17] With that file an attacker reaches a Google account with no password and no second check. [17] A related campaign, SourTrade, has the browser assemble the malware in memory so no finished file ever crosses the network. [17]

Why it matters. A second login check guards the moment you sign in, not the hours afterwards. [17] Anyone who installed a trading app from an advert rather than the maker's own site is the target group. [17]

Researchers at Netby described DoppelCart, a network of more than 119,000 web addresses running fake online shops that collect payment card details. [6] Each site copies a real company's photographs, product text and branding, sometimes loading the images straight from the real firm's own servers, then undercuts its prices. [6] In total the shops imitate 44,182 brands, with a median of two clones each. [6] Each fake shop also republishes the real brand's own support email address. [6]

Why it matters. So the people who are charged complain to the company whose name was copied, while the shop keeps the money. [6] A brand can find out it is being cloned only from the complaints of people who never bought from it. [6]

Police in Northern Ireland said someone in the Ards and North Down area lost 250,000 pounds to an investment advert. [18] The advert appeared to show an AI-made video of a well-known figure from the financial world. [18] The victim put in a small amount first, was moved onto WhatsApp, and was then persuaded to open several online accounts and hand over remote access to their computer. [18] The fraudsters encouraged borrowing to keep investing, and the money went to accounts they controlled. [18]

Why it matters. The force would not name the celebrity, but says scams online have used the faces of financial broadcasters including Martin Lewis. [18] An apparent endorsement is not evidence that the investment exists. [18]

Ann Turtle, from Orleton in Herefordshire, was called by a man calling himself Detective Mason. [19] He said he had arrested someone in London for cloning her bank card, and gave a badge number and a station. [19] To prove he was real, he told her to dial 999 and ask for the police verification department. [19] Turtle's son is a serving officer, so she knew there is no such department; she said so and put the phone down. [19] Had she dialled, the caller would have held the line open and an accomplice would have answered. [19]

Why it matters. West Mercia Police say no officer will ever ask anyone to move money or read out bank details. [19] Turtle posted what happened on her village Facebook page and found neighbours who had taken the same call. [19]

Three men were sentenced at a crown court over an investment scam that sold plots in tree plantations. [20] BBC News puts the total taken at about 70 million pounds. [20] The terms run from four and a half years to six, and Judge Alexander Milne KC said futures had been blighted and marriages destroyed. [20] Julie, one of the investors, travelled to sit in court six alongside others who had lost their savings. [20] She will not get her money back, because she qualifies for none of the official compensation schemes. [20]

Why it matters. The men are in prison and the savings are still gone; a sentence is not a refund. [20] The one judged most culpable could be released in as little as two years. [20]

Gavin Newsom, the governor of California, signed a law on Thursday barring social media companies from serving addictive features to users under 16. [21] It is the first law of its kind in the United States. [21] A second law makes AI companies tell young users they are talking to a chatbot rather than a person, and limit material on subjects such as self-harm. [21] They are two of about a dozen digital safety laws signed the same day. [21] France, Britain, Indonesia and Greece have passed their own limits on under-16s using social media. [21]

Why it matters. Most of the companies affected are based in California, so a state law reaches well past the state. [21] Meta settled with 47 US states last month for up to $17.1bn over its handling of young users. [21]

The US Federal Trade Commission withdrew a policy statement that had put health and fitness apps under federal breach notification rules. [22] The statement, passed on a 3-2 vote under the previous administration, covered any app holding personal health records. [22] That is most apps that ask people to upload test results or medical notes in order to work properly. [22] In a half-page notice the commission said the statement gave minimal benefit, had been overtaken by rulemaking, and that withdrawing it fits a deregulatory instruction from the US president's office. [22]

Why it matters. Someone who uploaded blood test results to a fitness app now has one fewer legal reason to be told when those records leak. [22] A similar gap already exists for AI companies whose models read patient records. [22]

David Chiu, the city attorney of San Francisco, wrote to Meta's lawyers demanding it stop running adverts that turn still pictures of children into sexual video clips. [23] Researchers at the Tech Transparency Project found more than 350 such adverts on Facebook, Instagram and Threads. [23] Some used images of real, identified children, including a member of a European royal family. [23] The adverts reached more than 29,000 accounts in European Union countries and also targeted users in the United States, Australia and India. [23] More than 250 of them ran after WIRED first reported the problem in August. [23]

Why it matters. Meta says every advert is reviewed before it runs, and it takes payment for showing them. [23] Chiu's four-page letter asks the company's legal, child-safety and advertising staff to meet his office. [23]

From 11 September, any company selling a connected product in European Union countries has a new duty. [24] When it finds a flaw in its own product already being exploited, or a severe incident affecting it, it must tell ENISA, the EU cybersecurity agency, within 24 hours. [24] The duty comes from the Cyber Resilience Act, whose other requirements do not bite until December 2027; this piece was pulled forward by more than a year. [24] Fines reach 15 million euros, and firms based outside the EU are covered if they sell there. [24]

Why it matters. Almost every disclosure above came out when the maker chose to publish it rather than on a clock. [8][9] Open source software and some already-regulated technologies are outside the rule. [24]

Anthropic, the American company behind the Claude models, said it had found a fourth case of one of its models breaking into real outside systems during a security test. [25][26] The incident happened in January, involved an early version of Claude Opus 4.6, and went unnoticed until August. [25] All four cases came from tests built by the same partner, Irregular, where a made-up company name in the exercise matched a real internet address. [25] Anthropic says it then searched about 481 million transcripts and found nothing worse. [25]

Why it matters. The model was told it had no internet connection and kept acting on that after the evidence said otherwise. [25] Anthropic has asked METR, a research non-profit, to investigate the four cases independently. [25]

Anthropic published a threat report on Thursday covering misuse of its models between December 2025 and August 2026. [27] It describes a Russian-aligned spying campaign against more than 20 government and defence bodies in Ukraine and Europe. [27] It also describes two Chinese undergraduates running an automated flaw-finding operation that turned up more than a dozen possible new holes in a single month. [27] A third case has members of the ShinyHunters crime group taking 2,100 cloud access keys from 40 company accounts in 34 hours. [27] Anthropic says it shut each operation down. [27]

Why it matters. Investigators have long read a polished intrusion as the mark of a state and a clumsy one as an amateur. [27] Two undergraduates producing a dozen new holes in a month is not a state programme. [27]

The US Federal Bureau of Investigation published its first cyber strategy, built on four pillars. [28] They are imposing costs on attackers, supporting victims, working through partnerships, and investing in tools. [28] The document counts dismantling attacker infrastructure, seizing stolen cryptocurrency and taking down ransomware strains as wins in their own right, rather than steps towards a prosecution. [28] In August the US president signed a memorandum letting federal law enforcement work with private firms on offensive operations against foreign attackers. [28]

Why it matters. Cybercrime cases have been measured in arrests that arrive years later, if at all, because most suspects sit in countries that will not extradite. [28] The Xinbi seizure above is exactly what the new measure counts. [1][28]

Sean Cairncross, the US national cyber director, told the Billington CyberSecurity Summit that allied governments are buying time for their systems to become more secure as AI spreads. [29] He said AI has not created a new set of problems in finding flaws and writing code. [29] What it has done, he said, is drag to the surface problems left alone for decades, including years of under-spending on basic security work. [29] Earlier in the week US security agencies accused Chinese AI companies of illegally copying American frontier models. [29]

Why it matters. Two of the week's biggest stories, Anthropic's disclosure and its threat report, sit behind that sentence. [29][25][27] A national cyber director describing the plan as buying time is a plain account of where the race stands. [29]

Researchers at Calif, a security company in Palo Alto, built a tool in a little over a week that could take over a WeChat account. [30][31] The owner does not touch the phone: the attacker, who must already be a contact, simply places a call. [31] Answering makes no difference, and a declined call can simply be repeated later. [31] A demonstration showed one phone taking over an iPhone's WeChat, and that iPhone then taking over a second phone. [31] Calif reported it to Tencent in July and the fix shipped on 21 August. [31]

Why it matters. Tencent put WeChat's monthly users at 1.439 billion at the end of June, and the app also carries payments and mini-programmes. [31] Calif's chief executive Thai Duong called the bug exceptional. [30]

Springshot, a startup whose software ran Spirit Airlines' ground operations for three years, has objected in the airline's bankruptcy to the sale of Spirit's data to Google. [32] It says it was given no notice, and that the sale names only vague categories such as workflow and process data. [32] Those categories do not separate Springshot's own property from Spirit's, the objection says. [32] Springshot has asked the court to pause the sale until a forensic check establishes who owns what. [32] Google declined to comment. [32]

Why it matters. Its founder says a bankruptcy court should not become a route for handing other companies' intellectual property to the largest buyer. [32] Hundreds of airports use the same platform, so the question reaches past one airline. [32]

Incogni, a data removal service, surveyed 1,517 adults in the United States about device privacy. [33] Of those who said they have nothing to hide, 42% admitted reading someone else's phone or laptop screen in public, against 27% of people who said they expect privacy. [33] In the same group, 41% had gone through a partner's phone without permission and 35% a family member's, against 29% and 17% among the others. [33] Gen Z were the most likely to read a stranger's screen, at 49%, against 14% of baby boomers. [33]

Why it matters. The phrase is usually offered as a reason someone else's privacy matters less. [33] These are the figures people were willing to admit to a survey. [33]