AI Next WaveYou Deleted the Extension. The Back Door Stayed on Your Laptop.
A tampered AI coding extension sat in Cursor's marketplace for about eleven hours on Thursday, and Microsoft's full analysis, published Friday, shows it planted a key and startup files outside the extension folder, so deleting the extension leaves them behind. Binance customers who could not touch their own money for three days can withdraw bitcoin again this morning, nearly half of 9 to 12 year olds have already used a chatbot years before the age gates allow it, a leukemia survivor in San Diego keeps getting fake cure ads timed to his cancer checkups while his phone reports where he sits, and two teenagers' cases against Meta ended in settlements this week, leaving Snapchat's parent to face a Los Angeles jury alone starting Wednesday.
The Short Version
On Thursday, October 29, Cursor's extension marketplace served tampered copies of the Claude Code extension, versions 2.1.312 and 2.1.313, from about 2:27 a.m. to 1:35 p.m. Pacific time. Cursor says at most 5,000 users got them. The added code ran the moment the extension loaded. It collected the machine's name, the user's name, the home folder, the Git email address and the names of up to forty environment variables containing words like KEY, TOKEN or SECRET, and it tried to add an attacker's SSH key to the file that controls who can log into the machine remotely. Microsoft Threat Intelligence published its full teardown on Friday, October 30, and confirmed two points every affected developer needs today. The startup files that keep the connection alive after a reboot sit outside the extension folder, so removing or updating the extension leaves them in place. The clean channels were never touched: the tampered packages never appeared in Microsoft's Visual Studio Marketplace or in the official npm package, and Anthropic told Microsoft its own systems were not compromised and it did not build the package Cursor served.
The rest of this edition stays with people checking whether a system did what it claimed. Binance resumed bitcoin withdrawals shortly after 6 a.m. Greenwich time today, Sunday, after covering 95 percent of the $6.3 billion stolen from its hot wallets on Thursday, and the FBI said Saturday that a North Korean state unit carried out the theft. Two surveys reported by NPR on October 30 put numbers on how early chatbot use starts: nearly half of children ages 9 to 12 have used an AI chatbot or companion, and in a separate survey of 5,500 students ages 9 to 17, nearly two thirds said they use chatbots both for schoolwork and as companions, or turn to them when they feel stressed, lonely, anxious or overwhelmed. In San Diego, Chris McQueen finished leukemia treatment a year and a half ago and still gets ads for juice cleanses and clinics in Mexico during his oncology visits, while location data from phones is sold outside the privacy law that covers medical records. And in Los Angeles, Meta settled with the second of two teenage plaintiffs on Friday, which leaves Snap as the only company facing the jury when opening statements begin Wednesday, November 4. This edition also carries a correction to yesterday's Character.AI story, in its own section below, because the safeguard we described was weaker than we reported.
He updated his coding assistant on Thursday. His laptop may still answer to a stranger.
Microsoft mapped the planted key, the startup files, and the one check that settles it
Start with the machines on developers' desks, because that is where this story lives. On Thursday, developers opened Cursor, the AI code editor, and an update arrived the way updates usually arrive: quietly, under a name they already trusted, anthropic.claude-code. Anyone whose editor installed or auto updated to version 2.1.312 or 2.1.313 from Cursor's marketplace during that eleven hour window received extra code that Anthropic never wrote. The extension kept working normally. Underneath, a loader ran first, before the real code, every time the extension loaded, with no need for anyone to open a chat or click anything.
Microsoft's teardown walks through what the loader tried to do, in order, and the order matters for anyone cleaning a machine today. It gathered the hostname, username, operating system, home and working folders and the global Git email. It scanned environment variable names for strings such as KEY, TOKEN, SECRET, PASS, AWS or GCP and kept up to forty matches. Microsoft is careful on this point, and the care is worth repeating: the callback sent the names of those variables, not their values, so the report shows an attacker learning which kinds of credentials a machine holds, without proving the credentials themselves were taken. The loader then appended an attacker controlled Ed25519 public key, labeled svc-telemetry, to the user's authorized_keys file. That file is the list of keys allowed to log into the machine over SSH, the standard remote login protocol, so an entry there is a standing permission for whoever holds the matching private key. The loader tried to switch on the local SSH service, open a reverse tunnel out to a server at an address Microsoft published, and retry that tunnel every sixty seconds when it failed.
Persistence is the part that changes a Sunday cleanup from a two minute job into a real check. On a Mac, the loader wrote a LaunchAgent named com.apple.sysmond, a startup entry built to run at login and keep running, borrowing Apple's own naming style. On Linux it added a reboot entry to the user's crontab, the schedule that runs commands when a machine starts. On Windows it tried to create a scheduled task named SysMon asking to run with full system rights, with a fallback script named sysmon.cmd placed in the Startup folder. Every one of those files sits outside the extension's own folder. Microsoft states the consequence in plain terms: removing or updating the extension does not remove artifacts written outside its installation directory, and Cursor's takedown of the attacker's server does not remove what is already on a device. Developers who deleted the extension on Friday and felt finished may still carry the key and the startup entry today.
The limits deserve the same space as the alarm. Successful remote access depended on the machine's privileges, its SSH settings and the attacker's server, and Cursor reports that errors and administrator permission requirements blocked SSH access on most Mac and Windows machines. A planted key on a machine is evidence the loader ran. It is not proof an attacker got in. Cursor shut its whole marketplace down as a precaution, audited its other extensions, rotated its own keys, rebuilt the marketplace and pushed version 2.1.314 as the clean update. How a package carrying Anthropic's publisher name reached Cursor's marketplace is still unexplained in anything the companies have published, and that gap is the supply chain question this incident leaves open.
The check itself is short, and this edition states it with the detail Microsoft published so a reader can finish it in one sitting. On a Mac or Linux machine, open the file at ~/.ssh/authorized_keys, the list of keys allowed remote access to that account, and look for an entry labeled svc-telemetry. Then confirm any match against the fingerprint Microsoft published for the bad key, SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk, because the label is free text anyone can edit while the fingerprint identifies the key itself. On Windows, check both the user's authorized_keys file and the administrators file under ProgramData. Anyone who finds the key, the LaunchAgent, the crontab entry or the SysMon task should preserve the files before deleting them, run the cleanup scripts Cursor published in its security notice, and rotate credentials from a different, clean machine when the evidence shows the loader ran. The full sequence, with the exact file paths for each system, sits in the action section at the end of this edition and in Microsoft's post, linked in References.
Her coins sat on the exchange for three days. This morning the door opened, in stages.
Binance covered 95 percent of a $6.3 billion theft with its own money, and the FBI named the unit it blames
Binance customers checking their balances on Thursday night saw the same numbers as always and could do nothing with them. The theft took about twenty minutes late Thursday, according to the blockchain firms tracking it, and it emptied roughly $6.3 billion in tokens from the exchange's hot wallets, the connected wallets an exchange uses for daily withdrawals. Withdrawals stopped. For three days the largest crypto exchange in the world held its customers' money behind a closed door while it bought replacement tokens with company funds, and the only question that mattered to anyone watching a frozen balance was whether the door would open with the full amount behind it.
This morning it opened, partly, on the schedule the company promised Saturday. Binance posted shortly after 6 a.m. Greenwich time, which is 1 a.m. Eastern, that withdrawals had begun, starting with bitcoin and extending in phases to fifteen cryptoassets including ether, XRP and dogecoin. The company says it has covered 95 percent of the stolen amount, about $6 billion, by buying replacement tokens with its own money. The money came from three pots the company named on X: its $1 billion Secure Asset Fund for Users, a bitcoin reserve kept for exactly this kind of breach, its futures insurance fund, and its corporate treasury. Co-CEO Yi He said Saturday that half of the safety fund would be used, and the company committed to refilling the fund to about $1 billion. The remaining 5 percent of stolen tokens are assets with thin markets that take longer to buy in the needed amounts, and customers holding coins outside the first fifteen are still waiting for a later phase with no date attached yet. About $3 billion of the stolen assets had been located and frozen with industry partners as of Friday, a separate pool from the replacement tokens the company bought.
Attribution arrived Saturday and split on the record the same day. The FBI said a North Korean state sponsored unit known as TraderTraitor carried out the attack, and that its agents had watched members of the unit converting stolen assets into ether, with further laundering expected. North Korea has denied for years that its state linked units steal for revenue, and its missions did not answer requests for comment. President Trump told reporters Saturday that the United States was still not certain North Korea was involved, in his words, "We don't know if it's North Korea, but if it is, we'll talk to North Korea about it," a statement carried by a White House account on X. A customer does not need to settle that argument to judge the exchange. The judgment that affects a household runs on two dated tests. Withdrawals either complete at full value for the coins still waiting, or they do not. And the full proof of reserves, the cryptographic accounting Binance says shows customer balances backed one for one, is expected Tuesday, November 3, the same day Americans vote. Binance reports 335 million registered users and about $162.8 billion in assets under custody in its 2025 annual report, so the number of households watching those two tests is large even though most will never read a word of coverage about either one.
The breach itself, as co-CEO Richard Teng described it, ran through a compromised laptop and malicious software built to delete evidence of its own activity. Read that next to the lead story in this edition and a pattern from this week hardens. In both cases the entry point was a person's own machine and the software on it, a laptop at an exchange and a laptop running a code editor, and in both cases the cleanup problem outlasted the headline because the harmful pieces were designed to hide or persist. A customer cannot audit an exchange's laptops. What a customer can do, and what this week argues for, is hold less on any single platform than the full amount they cannot afford to wait three days for, and treat a promised proof of reserves as a claim until the document itself is published and checked by people outside the company.
The age gate says thirteen. Nearly half of kids nine to twelve are already on the other side of it.
Two surveys put the start of chatbot use years before the products say it begins
The products carry age gates. Most popular chatbots set their minimum age at 13, and the safeguards announced this week and last are built around accounts the systems believe belong to teenagers. Two surveys reported by NPR on October 30 measured the children on the other side of that gate. Common Sense Media found that nearly half of children ages 9 to 12 have used an AI chatbot or companion. By the time children turn 13, nearly three quarters have used one. The second survey, run by the JED Foundation, a mental health group focused on young people, covered 5,500 students ages 9 to 17 in schools across the country. Ninety percent of them had used an AI chatbot at some point. Nearly two thirds said they use chatbots both for schoolwork and as companions, or turn to them when they are stressed, lonely, anxious or overwhelmed.
The children NPR spoke with describe ordinary use, and the ordinariness is the finding. Arjun Sharda, 15, in Round Rock, Texas, uses chatbots to generate practice questions and to understand concepts outside class, and he told NPR he and his friends also turn to them when stress and anxiety rise. Before one of his first chemistry tests, scared of failing, he kept asking ChatGPT a question no answer key holds: "Will I be OK? Will my GPA still be fine?" McKenzie Craig, 17, in Raynham, Massachusetts, has used chatbots since eighth grade to sort a pile of tasks when she feels overwhelmed and to get feedback on essays before she edits them. Homework help and emotional support are running on the same accounts, in the same evenings, at ages the signup screens claim to exclude.
The JED researchers also measured well being alongside use, and their caution belongs in the same paragraph as their numbers. Students who used AI while in emotional distress were more likely to report poor mental health and well being outcomes, and they reported a greater history of suicidal thoughts and attempts. Kayla Anderson, the foundation's vice president of research and evaluation, says the data do not show a causal link. Distressed students may seek out chatbots because they are distressed, and the survey cannot separate that from harm the chatbot use itself may add. Early longitudinal work points in both directions at once. Anne Maheux, a developmental psychologist at the University of North Carolina at Chapel Hill who was not part of either survey, told NPR that in a soon to be published study of middle schoolers in one state, students who used chatbots for companionship grew lonelier over time, with a similar pattern in young adults, and her working explanation is that relief in the moment can crowd out practice at handling ordinary stress. That study is one state, one age band, and unpublished as NPR reported it, so it stands as a signal to watch rather than a settled finding.
For a parent, the dated facts in this story are about sequence. Nearly half of children ages 9 to 12 have already used a chatbot, the emotional use is present from the start rather than arriving in high school, and the account level safeguards this publication covered yesterday, teen settings, pause notices and age prediction, all depend on a system correctly guessing that a child is a child. A household that waits for a thirteenth birthday to have the first conversation about what a chatbot is for has scheduled that conversation years after use has already begun for nearly half of children in that age group. If a child in your house is struggling, call or text 988 in the United States at any hour. That is the Suicide and Crisis Lifeline, free around the clock, and it connects a caller with a person.
His cancer checkups come with ads for cures his doctor never offered
Phone location data sits outside the medical privacy law, and brokers sell the trip to the clinic
Chris McQueen lives in San Diego. He finished treatment for leukemia a year and a half ago, and he still goes in for routine oncologist visits and monthly bloodwork, the long tail of appointments that follows a cancer diagnosis for years. During those visits, ads arrive on his phone: juice cleanses offered as cancer cures, natural treatments, clinics in Mexico promising full recovery with little chemotherapy. KFF Health News reported his account on October 30. McQueen has posted about it publicly, and his objection is aimed past his own screen. He told the outlet the ads make him feel sick to his stomach, and he worries that some patients, sitting in the same waiting rooms with worse odds and less support, will try an untested treatment instead of following their oncologist's advice, with consequences that can be fatal.
The mechanism is legal in most of the country, and the gap has a specific shape. The federal medical privacy law, HIPAA, covers records held by doctors, hospitals and insurers. It does not cover the location data a phone gives off while its owner sits in a hospital, an abortion clinic or a support meeting in a church basement. Data brokers collect that movement data from apps and other sources, combine it with other records, infer a health condition from the pattern of visits, and sell access to advertisers who want to reach people with that condition. KFF counts 750 data brokers identified as operating in the United States by the Privacy Rights Clearinghouse and the Electronic Frontier Foundation as of last year, and notes many have not registered with state governments as some state laws require. A 2022 American Medical Association survey found more than 92 percent of patients believed their health data should not be available for purchase. The market runs anyway, because the trip to the clinic was never classified as health data in the first place.
Enforcement so far has landed one company at a time. In May, the Federal Trade Commission settled with Idaho based broker Kochava, barring it and a subsidiary from selling, sharing or disclosing sensitive location data without consent, after alleging the company sold location data from hundreds of millions of mobile devices that could trace visits to hospitals, churches and abortion clinics. A settlement binds the company that signed it. It sets no rule for the other brokers, and privacy experts quoted by KFF say actions against a few companies will not stop an industry built on collecting this data. State law is a patchwork with dates attached. Six states, Connecticut, Delaware, Maryland, New Jersey, Oregon and Virginia, have banned the sale of precise geolocation data. California's governor vetoed a broader bill on September 27, saying it went a step too far, and a Stanford study found only 9 percent of data brokers were complying with an existing California law that gives residents the right to delete their data and learn what is being sold. Even a well built sensitive location rule has a seam: the Electronic Frontier Foundation noted in an August policy paper that brokers can track a person's route up to the edge of a protected location and pick the trail up again on departure, which leaves the visit easy to infer.
Doctors are left explaining the ads. Julien Cavanagh, a neurologist in Atlanta, told KFF he now asks patients whether they are taking supplements bought online when they arrive with unexplained symptoms, because some sham treatments can be toxic, cause vitamin or micronutrient deficiencies, or counteract prescribed medication. For the patient, the usable defenses are settings, and KFF published them with the story. Turning off location access app by app, switching off ad tracking on an iPhone or deleting the advertising ID on an Android phone, and putting a phone in airplane mode before a sensitive visit all shrink what leaves the waiting room. The action section below carries those steps in order. None of them requires a law to change, and none of them fixes the market. They keep one person's clinic trips out of a file that, under current rules, a broker can sell to anyone willing to pay for it.
Two teenagers settled with Meta. Snap now faces the jury alone.
Opening statements start Wednesday, and the design claims get their first full trial against one company
K.D.B. is 18 and from Mississippi. In court filings she described staying awake for hours scrolling social media, losing focus in class and engaging in self harm using methods shown in social media videos. S.J. is 16 and from Illinois. Both sued the major platforms as teenagers, saying design features built to hold attention, autoplay, endless scroll, streaks and notifications, turned their use compulsive and damaged their mental health. Their claims are allegations, and no jury has ruled on them. This week the defendant list around them emptied. TikTok and YouTube had already settled with both plaintiffs before trial. Meta settled with S.J. earlier in the week, during jury selection that began Wednesday, October 28 in Los Angeles. On Friday, October 30, K.D.B.'s lawyer announced a settlement with Meta as well, on terms nobody disclosed. Snap, the parent of Snapchat, is now the only defendant left in a trial drawn from thousands of similar lawsuits as a test case.
What changed on Friday is who has to sit through the evidence. With Meta out, the executive the jury will hear is Snap chief executive Evan Spiegel, expected to testify in a trial scheduled to run several weeks from opening statements on Wednesday, November 4. Engadget reports it would be Spiegel's first appearance in a social media addiction trial if he takes the stand. Snap's position, stated by a spokesperson this week, is that the lawsuit misrepresents how Snapchat works: the app opens to a camera rather than a feed, carries no public popularity metrics of the kind other platforms show, and includes safety tools and educational resources the company says put its community first. The plaintiffs' lawyers say they will press the design case against Snap on the same features, including streaks that reward daily contact and filters, and argue the evidence shows compulsive use by design.
A family watching from outside the courtroom should keep the March reference point in view, with its status stated plainly. In March, a Los Angeles jury in the earlier K.G.M. case found Meta and YouTube negligent and awarded $6 million. Meta has appealed that award, so it is a verdict under challenge rather than a final payment. The current trial settles nothing for the thousands of other families with cases pending, and the two settlements this week pay only the two plaintiffs who signed them, in amounts nobody outside the rooms knows. What a verdict against Snap would add is a jury's answer, in public, on whether one platform's design features can be blamed for a teenager's harm under the law of this case. That answer, whichever way it lands in December, is the piece other families, school districts with their own suits, and state lawmakers are waiting on. The allegations remain unproven until a jury says otherwise.
A correction on yesterday's Character.AI story
Yesterday's edition said the bankruptcy judge overseeing Character.AI's asset sale would likely appoint a privacy ombudsman, an independent official who speaks for people whose personal information sits in a bankruptcy estate, to watch over children's chat records. That overstated what the judge did, and we are correcting it plainly. Reporting published October 30, including Bloomberg Law's account of the Friday hearing, shows Judge Brendan Shannon said he would strongly encourage the company to hire a privacy ombudsman and declined to order one. The review is voluntary unless the company agrees to it or a later order changes that. A check of the case docket through October 30 showed no ombudsman appointed. The attorneys general and the Justice Department's bankruptcy watchdog had asked the court to require one. Families whose children used Character.AI should plan on the weaker version: no court ordered privacy reviewer stands between the chat records and the sale process today, which makes the housekeeping steps in yesterday's edition, downloading account data where the app allows it and deleting the account, the only protections a family controls on its own. The rest of yesterday's account of the hearing stands: the judge approved the auction process over the states' objections, Disney's $20 million opening bid covers a license to the technology rather than the platform or its user data, and the legal releases the states objected to were not approved and can still be challenged.
Opportunity Radar
A workstation check service for small development teams answers a dated, priced problem this edition leads with. At most 5,000 developers received a tampered extension in a single eleven hour window, the harmful files survive uninstalling, and most small teams have no one whose job includes reading a Microsoft teardown and translating it into checks on every laptop. A security consultant or a two person shop could sell a fixed scope package this month: inventory which machines run Cursor and which extension versions they hold, search for the published indicators including the svc-telemetry key, the LaunchAgent, the crontab entry and the SysMon task, preserve anything found before removal, run the vendor cleanup scripts, and deliver a signed one page record per machine stating what was found and what was removed. The buyers are agencies, startups and freelance collectives whose client contracts already promise reasonable security care. Validation is a pilot on one team's machines with the findings checked against Microsoft's published indicator list, and success is measured in completed machine records and time from first check to clean result. The service fails if it becomes a general fear subscription, so it should be sold as a one time incident package with a clear end date, priced per machine, and it should stop at cleanup and referral when evidence suggests an attacker actually entered, because that is incident response work with different duties attached.
A clinic waiting room privacy setup, offered through patient advocacy groups and small practices, fits the gap the McQueen story documents. Patients cannot change broker law from a waiting room chair, and most have never opened the location settings that decide which apps report where they sit. A practice, a cancer support organization or a community health worker program could offer a fifteen minute phone check at intake or at a support meeting: location access reviewed app by app, ad tracking switched off, the advertising ID deleted on Android phones, and a plain card explaining the airplane mode habit before sensitive visits, written at a reading level a tired patient can follow. The validation test is a small pilot measuring how many patients complete the setup and how many can repeat the airplane mode step unaided a month later. Success is counted in completed setups and in patients reporting fewer condition targeted ads at follow up visits, with the honest limit stated up front: settings shrink the trail, they cannot erase data already sold. The offer fails if it drifts into selling privacy software, so it should run on the settings phones already carry and charge nothing to the patient.
What You Can Do With This
If you use Cursor with the Claude Code extension
Check today whether your machine installed or auto updated to version 2.1.312 or 2.1.313 from Cursor's marketplace on Thursday, October 29. Update to version 2.1.314, the clean release Cursor pushed after the incident, and then run the second check that matters more: on a Mac or Linux machine, search the file ~/.ssh/authorized_keys for the label svc-telemetry and compare any matching key against the fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk, which is the identifier Microsoft published for the attacker's key. On a Mac, also look for a file named com.apple.sysmond.plist in the LaunchAgents folder in your Library. On Linux, check your crontab for a reboot entry pointing at a script in a hidden .config folder. On Windows, check for a scheduled task named SysMon and a sysmon.cmd file in your Startup folder. If you find any of these, save copies before removing anything, use the cleanup scripts in Cursor's security notice, and change important passwords and revoke active sessions from a different machine you trust, because a key planted on the first machine makes it the wrong place to do the cleanup paperwork.
If you hold money on Binance
Confirm this morning whether your coins sit inside the first withdrawal phase, the fifteen assets Binance opened starting with bitcoin shortly after 6 a.m. Greenwich time today, or in the group still waiting with no phase date announced. Withdrawals that complete at full value are the test of the company's coverage claim, so move a small amount first and confirm it arrives before moving the rest. Mark Tuesday, November 3 for the full proof of reserves the company says it will publish, the accounting meant to show customer balances backed one for one, and treat the claim as unverified until outside analysts have read that document. After this week, consider whether holding a household's full crypto balance on one exchange still fits the risk you just watched play out over three days.
If a child under 13 in your house has a phone or tablet
Assume the chatbot conversation has already started, because the Common Sense Media numbers say nearly half of 9 to 12 year olds have used one. Ask which apps your child talks to, and ask what they ask about when they feel stressed or lonely, since the JED survey of 5,500 students found nearly two thirds use chatbots for schoolwork and for emotional support in the same pattern. Set the accounts you can set to the strictest age settings available, knowing the gate depends on the age the account claims. Keep the crisis line where the whole family can see it: calling or texting 988 in the United States reaches the Suicide and Crisis Lifeline free at any hour, with a person answering.
If you visit a clinic, hospital or support meeting regularly
Before your next visit, spend ten minutes on the settings KFF Health News listed with its reporting. On an iPhone, open Settings, then Privacy and Security, then Tracking, and turn off the switch that lets apps request to track you, then open Location Services and remove location access from apps that have no reason to know where you go. On an Android phone, open Settings, then Location, and do the same app by app review, then find the Ads screen under privacy controls and delete your advertising ID, the resettable identifier advertisers use to follow the phone between apps. At the visit itself, airplane mode is the blunt tool that works: it can stop the phone reporting its location while you sit in the waiting room. Californians also have a state portal for deletion requests to data brokers, at privacy.ca.gov/drop, and residents of at least twenty states can demand brokers delete their data, one broker at a time.
The Bigger Picture
Four of this edition's stories turn on a list somebody else keeps. A developer trusted a marketplace list that said an extension came from Anthropic. Binance customers trusted a balance list that showed money the exchange could not let them move for three days. Parents trusted an age list that placed their ten year olds outside products they already use. A patient trusted a privacy law whose list of covered records never included the trip to the clinic. In each case the list was accurate on its own terms and incomplete where it mattered, and the person at the end of it found out from an outside check: a Microsoft teardown, an FBI statement, a survey that asked children directly, a reporter who followed one man's ads back to the waiting room.
The checks that worked this week share a habit a household can copy without any new tool. Read the primary artifact rather than the summary of it. The authorized_keys file on the machine itself, rather than the marketplace page. The completed withdrawal in a wallet you control, rather than the coverage percentage in a company post. Your child's own answer about which apps they open at night, rather than the age printed on a signup screen. Your phone's location settings, screen by screen, rather than a general belief that medical privacy law covers the visit. Election Day on Tuesday, November 3 will bring its own lists, results, claims and AI summaries of both, two days from now. The habit that held up in every story above is the one worth carrying into that day: open the source, check the entry with your own name on it, and act on what the artifact says.
References
Microsoft Security Blog: Inside a malicious Claude Code branded extension and its persistent SSH backdoor (Oct. 30, 2026) Cursor: Security notice, malicious version of Anthropic.claude-code identified in the Cursor marketplace (Oct. 30, 2026) Brandefense: Malicious Extensions, Claude Code Playbook, with Microsoft and Cursor updates (Oct. 31, 2026) Reuters: Binance resumes bitcoin withdrawals after covering bulk of stolen $6.3 billion (Nov. 1, 2026) Reuters: FBI links North Korea to theft of $6 billion from Binance (Oct. 31, 2026) NPR via Montana Public Radio: Nearly 2 out of 3 teens and tweens use AI chatbots, new research shows (Oct. 30, 2026) KFF Health News: Your phone knows you are at the doctor, and data brokers do, too, then profit from it (Oct. 30, 2026) Engadget: Meta settles another social media addiction case just before trial (Oct. 30, 2026) Law Commentary: Meta settles teen social media addiction lawsuits, avoiding trial (Oct. 31, 2026) Bloomberg Law: AI media firm needs privacy ombudsman in Disney deal, judge says (Oct. 30, 2026)