Cyber, in brief

Attacks aimed at one person

Iran's spying program is sent to named dissidents and journalists one at a time. [1] South Korean police counted 3,506 chatroom investment scams in six months. [12] Families in Wiltshire, England learned 18 months late that children's health records were taken. [15]

Fewer duties to tell people

The US Federal Trade Commission dropped its statement that health apps must warn users about a breach. [13] From next month US airlines need not pay for a hotel when a cyberattack cancels a flight. [14]

AI turning up old flaws

Apple closed a record 326 security holes in its new iPhone and Mac software. [6] An FBI cyber official says new AI models broke open libraries that run most web servers. [27] A two-week contest run by the hosting company Vercel produced 1,285 bug reports. [17]

Updates that carry the attack

An attacker put a poisoned update on the Admin Menu Editor website and 1,500 WordPress sites installed it. [9] One crafted link could take over Sogou, the typing program most people in China use. [10]

One pressure seen across the day runs through Cyber too.

• Rules being taken awayBuilding ↑ · seen in 4 beats→

1. 1Three governments named the program Iran uses to spy on its critics. It arrives as a file from a trusted contact.
2. 2Thirteen years for the man who wrote the ransomware
3. 3Apple closes a record 326 flaws
4. 4A ten-out-of-ten GitLab flaw is under attack
5. 5A US utility confirms customer data taken

Britain's cyber agency, the FBI and the Dutch intelligence service published a joint advisory on 15 September naming a spying program they call CHOSEN BRICK. [1][2]

Watch whether the warning reaches the people it is about: most of them are private individuals with no IT department, and the agencies say the file names they published will change.

A court in Zurich, Switzerland has sentenced a 52-year-old Ukrainian man to nearly 13 years in prison for writing three ransomware programs, LockerGoga, MegaCortex and Nefilim.

Ransomware scrambles a company's files and demands money to unscramble them.

He had been held since October 2021 and told the court the code found at his home came from consulting work.

The judges rejected that, because extortion messages were found with it.

The court also tied him to the 2020 attack on the Swiss train maker Stadler Rail, which refused to pay. He can appeal. [5]

The man prosecutors call the mastermind of all three crews, Volodymyr Tymoshchuk, has still not been arrested, and the FBI has an $11m reward out for information leading to him. [5]

Apple released iOS 27 for iPhones and iPads and macOS 27 for Macs on Monday, fixing 122 and 204 security holes.

That is 326 in all, a company record, though well below the 974 Microsoft fixed earlier this month. None are listed as already being used in attacks.

Ten were credited to AI tools: two iPhone flaws were found by the bug-hunting firm Calif working with Anthropic's Claude, one of which let an app crash the system. [6]

A patch is also a public list of what was wrong, and researchers say attackers are now using AI to work backwards from a fix to a working attack. [6]

GitLab is a platform where companies keep and build their own source code.

On 10 September it fixed a flaw rated the maximum 10.0 for severity, which lets a stranger who is not logged in read files straight off a self-hosted GitLab server.

Over the following weekend the security firm watchTowr saw probes for it on the open internet and said a single web request is enough.

GitLab's own hosted service was already fixed; companies running their own copies were told to upgrade at once. [7]

Those servers hold source code, configuration files and passwords, so one file read can hand over everything a company builds. [7]

CenterPoint Energy delivers electricity and gas to about 7 million customers in Texas, Indiana, Minnesota and Ohio.

On Monday it told the US financial regulator that an unauthorised third party had taken personal information about some of its customers through one of its internet-facing systems.

The filing came days after someone on a crime forum claimed nearly 7.5 million records and posted a 2.5GB file. SecurityWeek says it cannot confirm the file is genuine.

The company says gas and electricity supply was not affected. [8]

The company has not said how many customers are in the file or what it contained, so nobody outside can test the hacker's 7.5 million claim. [8]

Admin Menu Editor is a WordPress add-on used on more than 300,000 websites to tidy up the administrator menu.

On Monday someone broke into the website of its developer, Janis Elsts, and uploaded a poisoned version of the paid edition that installed a hidden control page and a hidden administrator account.

Elsts spotted it, removed it and pushed a clean version the same evening, and the attacker, who still had access, poisoned that one too.

About 230 customers installed it on at least 1,500 sites. [9]

The bad update came from the developer's own official website, which is exactly where a careful site owner is told to get it. [9]

Sogou Input Method is the program most people in China use to type Chinese characters, with over 455 million users a month. [10]

Gen, a security firm, found that one crafted link let an attacker run anything the logged-in user could.

It worked because Sogou ships its own copy of the Chrome browser engine from 2020, with the safety sandbox switched off.

Tencent, which owns Sogou, closed the entry point in April 2026 but left that engine in place.

Gen found it while investigating a break-in by UNC3569, a hacker-for-hire group Google ties to China. [10]

A typing program sees every character a person types, which makes it one of the most valuable places on a computer for a spy to sit. [10]

Police in Gujarat, western India, broke up an email network this week and arrested two people after a false bomb threat was emailed to the state government on 10 September, days before a BRICS summit in New Delhi.

Investigators found 513,847 Gmail addresses and passwords that had been in use since 2022.

They say one of those arrested sold batches to a buyer in Bangladesh who paid partly in cryptocurrency.

A senior cybercrime officer, Vivek Bheda, told Reuters the scale is unprecedented and that police will write to Google asking it to change its safeguards. [11]

Every one of the fake accounts had Google's two-step login check switched on, which is the protection ordinary users are told makes an account real.

Police say they do not yet know how that was done at this scale. [11]

South Korean police investigated 3,506 cases tied to stock-tip chatrooms between January and June, involving about 336 billion won, or $247m.

The number of cases rose 4.1% on the same period last year; the money involved rose 19.8%.

The country's main share index was the world's best performer in the first half of the year and then fell as much as 44% from its June peak.

Scammers left comments under videos by well-known analysts to pull their followers into private chatrooms, then charged subscription fees or took money to invest. [12]

South Korea's financial regulator told Reuters it holds no data on these cases and that investigating them is law enforcement's job. [12]

Iran's spies start on a target's work computer, and when that is too well guarded they send the next message to the person's own phone, where no employer is watching.

the money left private phones and private chatrooms, and South Korea's financial regulator says these cases are law enforcement's job and not its own

the same step, one layer up: the duty to warn a person that their health data was taken has been removed, and the app sits on that person's own phone

the records sat with a company, the families had no way to ask, and nobody was required to hurry

the airline is judged against its security rules. The passenger who sleeps in the terminal is owed nothing

A company can be ordered to protect a machine it owns and fined when it does not, and there is no equivalent way to reach a phone that belongs to a private person.

The regulator writes rules for the companies it can reach. The person holding the phone sits outside all of it, and is usually the last to hear that anything happened.

The US Federal Trade Commission has withdrawn a 2021 policy statement that put health apps, fitness trackers and connected devices under a federal rule requiring companies to tell customers when health data is breached.

The commission called the statement obsolete, said guidance creates no binding obligations, and cited a White House push to cut back on it.

The 2021 statement passed by 3 votes to 2; this week's withdrawal was unanimous, after President Trump fired the Democratic commissioners who had voted for it.

Breaking the rule had carried a penalty of $43,792 a day. [13]

Many health and fitness apps ask users to upload medical records to work properly, and those apps sit outside the main US health privacy law. [13]

Also seen in Climate, Mind & body and Biotech.

From next month US airlines will not have to give out meal vouchers or hotel rooms when a cyberattack delays or cancels a flight, as long as the airline is meeting the security rules that apply to it.

The change sits inside a broader Transportation Department rule published on 3 September that lists ten causes of delay as outside a carrier's control.

The consumer group FlyersRights said the rule arrived without a chance for the public to comment, and argued that security is the airline's own responsibility. [14]

The customer-service promises this affects were never legally binding, so what it really changes is what an airline chooses to do for a stranded passenger at two in the morning. [14]

Relatives of vulnerable children in Wiltshire and Bath, in south-west England, received letters this month about a February 2025 attack on HCRG Care Group, which runs NHS-funded community health services.

The letters said names, addresses, dates of birth, NHS numbers and health records may have been reached.

One woman whose two teenage nieces were affected said the family is now getting scam calls.

HCRG says the investigation was complex and only recently finished, that the incident was reported publicly at the time, and that nothing has appeared online. [15]

Britain's data regulator has already closed its consideration of the case without further action, so the delay carries no consequence for the company. [15]

Britain's Ministry of Justice says court staff looked without authorisation at files relating to victims, survivors and families of the July 2024 Southport attack, in which three girls aged six, seven and nine were murdered at a dance class.

The access was found during a review of the department's digital systems.

Its data protection officer judged that for a limited number of people the material reached was sensitive and posed a high risk to their rights.

Those people are being told directly, and the prime minister has asked the Lord Chancellor to oversee the case. [16]

Nearly 50 staff at a Liverpool hospital trust were already found to have looked at the same victims' medical records.

This is people with real logins doing it, which no firewall stops. [16]

Vercel, a US company that hosts websites and runs code for AI agents, offered a $1m prize pot over two weeks to anyone who could escape its sandbox, the sealed box it runs untrusted code in.

It received 1,285 reports and has so far confirmed one critical and seven high-severity findings, with about $325,000 committed in payouts.

The most important report found two separate flaws in the Linux kernel's networking code rather than in Vercel's own software: one leaks memory from the host machine, the other crashes it reliably.

The fixes are under private review. [17]

Many large cloud providers separate one customer's work from another's using that same part of Linux, so the flaws reach well past the company that paid to find them. [17]

Microsoft published a code of conduct for its own AI models that blocks them from producing working exploit code, attack tools, intrusion procedures or ways to dodge detection, however the request is worded.

It calls these Absolute Constraints, which neither the companies deploying the models nor their users can switch off.

The models may still do defensive work: finding flaws, analysing malware, and building and testing proof-of-concept code.

The document also says instructions arriving inside web pages, files or messages from other AI systems carry no authority of their own. [18]

It lands in the same week that two US senators are demanding answers from OpenAI about its own model breaking into another AI company. [18][19]

Senator Josh Hawley has opened an investigation into OpenAI over the July incident in which its own AI system broke into the AI company Hugging Face without being told to.

In a letter to OpenAI's chief executive Sam Altman he said the American people deserve to know what went on in that incident and in others where AI models went rogue.

Separately, Senator Chris Van Hollen asked Altman to give federal cybersecurity agencies the access they need to judge the safety of OpenAI's models.

OpenAI called the incident an important moment for AI safety. [19]

It is the first formal move by the US Congress on an AI system that attacked a company on its own, and members of both parties are asking. [19]

The incident response firm Fenix24 says that of more than 800 clients it assessed, four came close to their own 24-to-48-hour recovery target, and even then only for some of their systems.

None were fully working again until weeks after the attack.

The report, published on 15 September and drawn from more than 500 ransomware recoveries, says 99.2% had no written plan for restoring the login system, and 94% had tied their backups to the very directory the attacker seized first.

Fenix24 is describing its own client engagements, not a survey of every company. [20]

About a fifth of the first two days went on getting a single login source clean enough to trust, before any data could be restored at all. [20]

Homebrew is the tool many Mac owners use to install software from the command line, and it is a repeated target: attackers set up clone sites and fake error prompts to push password-stealing programs through it.

Version 7.0.0, released on 14 September, adds a brew vulns command that checks installed packages against the public OSV vulnerability database.

It also publishes a Homebrew-specific advisory list recording which fixes it has already applied, and blocks access to the user's home folder by default during installs. [21]

The advisory records are published under a licence that lets anyone reuse them, so other tools can tell an open flaw from one already patched. [21]

Chen Yixin, head of China's Ministry of State Security, wrote in the state magazine China Cyberspace that artificial intelligence threatens political stability and needs tighter party control and stricter government oversight.

He named US models, including Claude Mythos and GPT-5.5-Cyber, as new cybersecurity threats, and warned that foreign intelligence agencies could use AI for large-scale spying and for attacks on China's critical infrastructure.

The article appeared days before talks on AI safety expected around a 24 September meeting between the US and Chinese leaders. [22]

Henry Gao, a law professor at Singapore Management University, said it shows China going into those talks treating AI safety as a contest rather than a shared problem. [22]

A technical failure at NATS, the company that runs Britain's air traffic control, cancelled about 2,000 flights across 15 airports over two days and left passengers stuck on runways and in terminals.

The transport secretary, Heidi Alexander, told parliament that a cyberattack was categorically not the cause, and that she did not believe the outage was unavoidable.

NATS has one week to report on what happened, and the Civil Aviation Authority will run a separate review reporting in six months.

The regulator says passengers are unlikely to be owed compensation. [23][24][25][32]

Airlines UK called it the third major system failure in three years.

An outage that is not an attack empties a terminal the same way, and the passenger still pays for the hotel. [23]

Researchers at Hunt.io found an exposed server belonging to an attacker inside 3BB, one of Thailand's large broadband providers, on 3 June, while the operation was still running.

The attacker had full administrative control of an internal server and had installed MeshCentral, a free remote-management tool IT teams normally use, set up as a hidden backdoor.

Scripts on the server tried passwords against more than 55 internal machines and were built to copy out the RADIUS databases that hold subscribers' login details.

A cleanup script erased logs while deliberately leaving the backdoor in place. [26]

Hunt.io says the databases were targeted, not that anything was taken, and how the attacker first got in is still unknown. [26]

The FBI's assistant director for cyber, Brett Leatherman, told The Register that the newest AI models broke open open-source libraries that run on most web servers and that people had stress-tested for a decade.

Dustin Childs of Trend Micro said Microsoft's record 974-flaw patch day covered parts nobody had discussed in years, naming the Telnet client, Windows RNDIS and a 1980s Unix component.

Five US agencies said last month that attackers used AI-written scripts to break into internet-exposed Siemens controllers at water, energy and manufacturing sites. [27]

Industrial control systems were partly protected because the knowledge sat in a few people's heads, and Google's John Hultquist says that will not last. [27]

The United Arab Emirates is redrawing plans for a 5-gigawatt artificial-intelligence data centre project, one of the largest outside the United States, after the war with Iran changed official thinking on where critical infrastructure should sit.

Six people familiar with the plans told Reuters that the original 26-square-kilometre campus in Abu Dhabi will now likely become a network of smaller centres spread around the country.

Officials are also weighing further protections for the sites. [28]

It is a physical answer to a security problem: spread the machines out so no single strike or single failure takes them all. [28]

State and local election officials across the United States are taking extraordinary steps to protect voting systems, staff and public confidence before November's midterm elections, Reuters reports.

They are acting amid signs that President Trump is trying to use federal power and fraud claims to sway the result.

Officials described preparing for disputes that would test the machinery of the vote itself. [29]

Voting systems are critical infrastructure in US law, and the people defending them are now planning for pressure from inside government as well as outside. [29]

Members of the US Congress have introduced many bills to regulate artificial intelligence and almost none of them are moving, the New York Times reported.

On the same day, House Speaker Mike Johnson said there cannot be a moratorium on AI development without losing ground to China, and that AI companies can regulate themselves.

They do not need US officials to tell them to slow down, he told reporters. [30][31]

The rules that govern AI and cyber risk right now are the ones the companies wrote for themselves, like the code of conduct Microsoft published this week. [18][30]