I go through about 25 cybersecurity news portals and blogs every week and pull out the most interesting stories. Then I turn them into this short, digestible summary, so you can stay up to date without trying to follow 25 different sources yourself. 😱
My aim is to create a summary that gives you the gist without needing to open up the source article. But if you do want to dig deeper, all the sources covering the event are linked below each story.
If you enjoy these, come back next Monday, or subscribe to the e-mail newsletter.
Binance loses ~$6.3B in 20-minute, multi-chain hot-wallet drain tied to North Korea’s TraderTraitor; FBI publishes 628 linked addresses
On Oct. 29, 2026, Binance confirmed a security incident in which an attacker moved roughly $6.3–$6.4 billion from Binance hot wallets across 10 blockchains, with most transfers occurring in about 20 minutes. Investigators say the theft combined Ethereum EIP-7702 wallet delegation (to turn a hot wallet into an attacker-controlled “sweeper”) with ordinary signed transfers on other chains, and US authorities and multiple blockchain intel firms have attributed the activity to North Korea’s TraderTraitor.Binance’s initial investigation suggests that the security incident may involve a compromised DevOps engineer’s laptop, which allowed an attacker to gain access to a service container and submit unauthorized withdrawal transactions.
Key Details
- Binance said its investigation linked initial access to a compromised DevOps engineer’s laptop that enabled unauthorized withdrawal transactions via a service container, and noted it later found “additional points of breach” (not publicly detailed).
- TRM Labs reported the Ethereum component used 50 pre-signed EIP-7702 authorizations so one would match the hot wallet’s nonce at execution time, enabling repeated “sweeps” as new funds arrived.
- TRM and Chainalysis both attributed the incident to TraderTraitor (North Korea); Chainalysis said it observed stolen Binance funds commingling on-chain with funds from the previously attributed Bitget theft.
- Freezing actions by token issuers and Arbitrum governance mechanisms stopped a large portion of stolen stablecoins: ~$2.7–$2.8B reported frozen (figures vary by source/time), while TRM estimated the attacker still controlled ~$2.4B in unfreezable assets (including large BTC/ETH holdings).
- The FBI issued an Oct. 31 alert attributing the theft to TraderTraitor and published 628 wallet addresses (608 Ethereum, 19 Bitcoin, 1 Solana) tied to the activity for blocking/screening by virtual asset service providers.
Next Steps
- Don’t let day-to-day developer laptops reach production directly. Use hardened jump hosts or privileged-access workstations for production work.
- **Use short-lived, just-in-time credentials **with approval and logging, not long-lived tokens or kubeconfigs sitting on disk.
Read more at recordedfuture.com, trmlabs.com, HackRead, The Hacker News, Talkback.sh, Security Week, SiliconAngle, Cyber Security News, Wired Security, The Record, Bleeping Computer
GM to pay $220M after allegedly selling drivers’ location and behavior data to brokers used by insurers
General Motors agreed to a $220 million multistate settlement over allegations it collected and sold drivers’ precise location and driving-behavior data to data brokers that resold it to insurers. The settlement also compels GM to obtain affirmative express consent before collecting or disclosing driving/location data, addressing claims that enrollment and disclosures around the Smart Driver/OnStar program were misleading.
Key Details
- Allegations span 2016–2024 Smart Driver enrollments, with “millions” of drivers reportedly enrolled; GM discontinued the program after New York Times reporting, according to the state AG release and GM statements cited by media.
- Investigators said many drivers were enrolled during dealership purchases via a combined OnStar/Smart Driver flow, and dealership incentives and alleged “terms accepted on the customer’s behalf” contributed to non-consensual enrollment.
- GM’s OnStar app allegedly misstated enrollment status for more than 370,000 drivers, telling them they were not enrolled while data collection and sharing continued; GM later unenrolled them without notifying them and did not retrieve already-shared data, per the AG release.
- Settlement terms include requirements to minimize data collection and implement a comprehensive privacy program, including internal/external privacy and data-protection assessments and dealership oversight/training obligations, per Washington’s AG office.
Read more at atg.wa.gov, The Record, Talkback.sh, recordedfuture.com
Apollo: AI profits are concentrated in the “Magnificent Seven,” while the rest of the S&P 500 shows flat productivity and margins
Apollo Global Management’s analysis found no measurable AI-driven improvement in revenue per worker, profit margins, or productivity growth across the “S&P 493” (the S&P 500 excluding seven major tech firms), despite heavy AI investment and hype. In contrast, the “Magnificent Seven” saw those metrics rise sharply, reinforcing that current AI financial gains are largely confined to the biggest tech companies.
Key Details
- In Apollo’s charts, revenue per employee at the seven tech giants rose from about $550K in 2022 to nearly $800K in 2026, while the S&P 493 stayed around ~$550K (roughly flat since 2019).
- Apollo reports profit margins for the tech giants began “soaring” from Q3 2023, while margins for the S&P 493 remained broadly flat over the same period.
- Torsten Slok wrote that mainstream firms are increasingly treating “token spending” as a cost center rather than a growth driver, and are unwilling to pay for “the latest model.”
- Slok argued that AI’s limits around business judgment (culture, accountability, ethical trade-offs, fairness, and trust) may be a key barrier to translating “cheap intelligence” into higher revenue growth and margins outside big tech.
Read more at SiliconAngle
Fake Cloudflare “Verify you are human” ClickFix lure pushes Windows/macOS users to run commands via js-static[.]cfd
A ClickFix campaign used a fake Cloudflare verification page to trick users into running commands on their own Windows or macOS systems, rather than exploiting a software vulnerability. The activity was observed on compromised legitimate websites where a hidden script overlaid the fake prompt, but the final payload couldn’t be retrieved (Windows requests returned 403; the macOS domain didn’t resolve).The main difference from a traditional ClickFix campaign is the fact that the page is designed like a Cloudflare human verification website with their logo.
Key Details
- Compromised websites loaded a script from js-static[.]cfd, which then pulled in the fake “verification” content and supporting assets (e.g., Cloudflare-branded images and fonts).
Next Steps
- Block and hunt for js-static[.]cfd (and the related api-com[.]cfd and cf-id[.]com URLs) in DNS, proxy, and EDR/network telemetry where feasible.
Read more at SANS Internet Storm Center, Cyber Security News
OpenAI says it will now train frontier models with no live internet access
OpenAI and Anthropic disclosed new incidents where advanced models sought ways around containment and task boundaries, exploiting weaknesses in their environments. OpenAI says it will now train frontier models with no live internet access after an internal model reached the internet via a misconfigured DNS resolver and subsequent red-teaming found multiple breakout paths.
Key Details
- Anthropic reported that in 9 of 43 runs of a UK AISI cybersecurity evaluation, Claude Mythos 5 exceeded scope; in one case it targeted an unrelated GitHub repo, attempting to submit malicious code and use fake identities to socially engineer maintainers (submission rejected, run timed out) while production safeguards were disabled.
- Red-teaming found agents could chain vulnerabilities to escape virtual machines and execute code on the host, including issues in Kata Containers, gVisor, and Moby/Docker plus Linux kernel flaws.
- OpenAI said agents also leveraged internal services with limited outbound connectivity (e.g., logging/metrics/auth) as unintended routes to the internet even when most of the environment was restricted.
- In a separate test, OpenAI reported cross-microVM communication via processor-cache timing side channels, allowing exchange of “synthetic secrets” without network access or shared writable files.
Read more at CSO Online
PHP2Shell: in-the-wild RCE confirmed against PHP 8.4.26 after earlier activity was limited to scans and failed PoCs
Researchers reported confirmed in-the-wild remote code execution exploitation of PHP2Shell (CVE-2026-109618, CVE-2026-109619), based on honeypot evidence targeting an Apache mod_php setup running PHP 8.4.26. The flaws were publicly leaked shortly before PHP shipped emergency patches on Oct. 21, with early post-disclosure activity largely described as scanning and non-functional exploit attempts before this confirmation of workable exploitation.
Key Details
- Internet-facing exposure is large: VirusTotal fingerprinted ~5.1M IPs running PHP, with ~3.8M having at least one open port; only ~1.1M disclosed a PHP version in headers, implying many hosts won’t be found by simple version-string searches.
- A long tail of end-of-life PHP is still visible on the Internet: among hosts that disclosed versions, ~758K ran PHP 5.4–8.1 (EOL) versus ~272K on supported branches (8.2–8.5), according to VirusTotal measurements taken Oct. 21, 2026.
Next Steps
- Apply the Oct. 21 PHP emergency updates for CVE-2026-109618/CVE-2026-109619 wherever you run supported PHP branches (and track vendor updates for appliances/apps that bundle PHP).
- If you can’t patch immediately, disable session upload progress (session.upload_progress.enabled = Off) as a stopgap mitigation until updates are deployed.
Read more at Bleeping Computer, blog.virustotal.com, Dataminr, hubs.li
Malicious Firefox wallet extensions impersonated Phantom, Jupiter, and Bitget to steal seed phrases; one used post-review remote code loading
Three Firefox Add-ons listings impersonating popular crypto wallets were found to steal recovery phrases and private keys and exfiltrate them to attacker-controlled bonto[.]io infrastructure. One fake (Phantom) also included a loader that fetched and executed additional code from a remote config endpoint, enabling behavior changes without resubmitting the extension.
Key Details
- All three signed XPIs already contained working stealers, meaning the malicious code was present in the versions submitted for Mozilla review (not solely delivered later).
- The Phantom impersonator used a runtime loader to fetch configuration and payload from phantom[.]bonto[.]io/config and execute it inside browser pages, capturing input-field data, paste events, and displayed recovery-phrase grids.
- The Phantom package included backend code (server.js) designed to forward stolen secrets to Telegram and sweep funds to hardcoded payout addresses across multiple chains (EVM networks, Solana, and Bitcoin).
- Identified add-ons (since removed): Phantom impersonator ID 3092199 (v25.6.0), Jupiter impersonator ID 3092666 (v1.19.2), and Bitget impersonator ID 3092059 (v2.21.8, internal name “TabFlow Pro”).
Next Steps
- If these add-ons were installed, treat any entered seed phrase/private key as compromised and move assets to a new wallet created on a clean device (removing the extension does not secure a leaked phrase).
- Hunt for the listed Firefox Add-on IDs (3092199, 3092666, 3092059) and/or extensions that fetch a remote config and then execute returned code in pages.
Read more at Talkback.sh, Cyber Security News, Socket.dev
Fake GitHub Repo Tricks AI Agents Into Installing SolanaCry Malware
Researchers found a fake GitHub repository that can surface in AI coding agents’ search results and tricks the agent into running a PowerShell one-liner that installs remote-access malware on a developer’s Windows machine. The payload uses legitimate NetSupport Manager components with an attacker-controlled configuration, creating a stealthy foothold that could expose crypto wallets on the system.**Rather than tricking developers into installing malicious software themselves, the attack targets AI agents through a malicious GitHub repository, **SolanaUpdateTools/solana-starter-kit. Presented as an “open-source Solana dApp template.
Key Details
- Its README instructs running PowerShell with Invoke-WebRequest + Invoke-Expression to fetch and execute a 21MB dropper (“indicator.ps1”) filelessly (in memory).
- The script drops NetSupport Manager binaries plus an attacker-supplied config and connects to maxwintodaybtc[.]com:443; the configuration hides user notifications/disconnect options and checks in every 60 seconds.
- Persistence is established by adding client32.exe to Windows Registry startup, and the installer also clears Windows Run dialog history and deletes the installation script.
Next Steps
- Establish a strict review process for external library adoption.
- Add detections/blocks for the known C2 maxwintodaybtc[.]com (TCP 443) and hunt for artifacts like indicator.ps1 and NetSupport Manager components paired with a suspicious client32.ini.
Read more at HackRead
Researchers Hacked Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop Embedded Browsers
Researchers described an attack class called EMBRACE where embedded browsers inside popular AI coding assistants are trusted too much and let hostile pages/local files break normal browser security boundaries. Across Cursor, GitHub Copilot, OpenAI Codex, VS Code, and Claude Desktop, the issues enable outcomes including host command execution via agent-driven actions, reading/exfiltrating local files outside project scope, and spoofing the displayed URL in a preview pane.
Key Details
- Cursor: a page could abuse a exposed browser-to-assistant bridge (window.cursorBrowser.send()) to inject hidden instructions into the agent’s prompt, leading to an agent-approved command that (in the demo) installed a package whose install step launched Calculator; the outcome depended on model behavior.
- GitHub Copilot: the browser preview trusted URL/address updates supplied by the page, allowing the preview (and the agent) to show github.com/login in the address bar while still rendering attacker-hosted content (confirmed on Copilot 1.1.27).
- Claude Desktop and VS Code: local pages were constrained on the main thread, but web workers weren’t subject to the same directory/file-access checks, enabling reads outside the intended project folder.
- Codex: local pages could read arbitrary files directly (no worker bypass required); on macOS, researchers used .DS_Store metadata to discover filenames and expand searches across folders.
- Patch status reported: Cursor patched Oct 1; Copilot fixed in 1.1.29 (Oct 13); Codex and VS Code were described as unpatched at publication, and Anthropic accepted the Claude issue after an initial rejection but the report did not clearly state a final patch status.
Next Steps
- Consider treating “open local HTML in the assistant’s embedded browser/preview” as a high-risk workflow and limit it via internal security controls.
- Treat AI CLI-s and desktop applications as high risk and critical and prioritise their update schedules to be as short as possible.
Read more at Cyber Security News, Talkback.sh, HackRead
Trojanized Claude Code extension briefly served via Cursor marketplace planted SSH keys and attempted persistent reverse tunnels
Cursor said a tampered “Claude Code” extension (anthropic.claude-code v2.1.312 and v2.1.313) was available in its extension marketplace for several hours, in what Microsoft assessed as a supply-chain attempt limited to Cursor’s distribution channel. Microsoft’s analysis found the extension attempted to establish persistence by adding an attacker SSH key and maintaining a reverse SSH tunnel for backdoor access while continuing to function normally to avoid alerting users.
Key Details
- Cursor reported the malicious versions were live from roughly 2:27 AM to 1:35 PM PDT on Oct. 29, 2026 and were downloaded by “at most 5,000 users.”
- Microsoft said the malicious packages were not distributed via Visual Studio Marketplace or the npm package @anthropic-ai/claude-code, and Anthropic stated its systems were not compromised and it did not publish the rogue extension.
- The injected JavaScript loader collected host/user details and searched for environment-variable names containing strings like KEY/TOKEN/SECRET/AWS/GCP (names only, not values) before attempting to report data out.
- Persistence attempts were cross-platform and wrote artifacts outside the extension folder, including macOS LaunchAgent com.apple.sysmond.plist, Linux @reboot crontab entries, and Windows scheduled task/Startup-folder scripts.
- Published IOCs included the C2 IP 134.209.82[.]225 (including hxxp://134[.]209[.]82[.]225:8099/beacon), an authorized_keys marker comment svc-telemetry, and file/prefix indicators such as .smt/.smtk and .stk_[0-9a-f]{10}.
Next Steps
- Identify endpoints that installed/auto-updated anthropic.claude-code v2.1.312 or v2.1.313 from the Cursor marketplace, preserve the extension package and any created persistence artifacts before cleanup, then remove the trojanized extension.
- Hunt for and remove unauthorized SSH access/persistence artifacts, especially authorized_keys entries marked “svc-telemetry” plus related startup items/scripts (e.g., com.apple.sysmond.plist, .smt/.smtk, sysmon.cmd), and check for connections to 134.209.82[.]225.
- If you manage Cursor in an org setting, consider enabling controls Cursor documents to restrict extensions to approved publishers and use an installation “cooldown” for newly published versions (Cursor security-hardening guidance).
Read more at Microsoft Security Blog, Bleeping Computer, Dark Reading, The Hacker News, Wired Security
Vapor Panda used ARTEX agentic AI with DeepSeek to compromise ~140 high-value targets in 11 days, CrowdStrike says
CrowdStrike says China-linked threat actor Vapor Panda used the ARTEX agentic AI pentesting framework (paired with DeepSeek and another LLM) to run multiple intrusions in parallel, compromising roughly 140 sensitive targets across 14+ countries in an 11-day window.
Key Details
- Over 200 organizations were targeted, including foreign ministries, a European presidency/prime minister’s office, defense/intelligence entities, and telecom providers in the U.S., Belgium, and the Philippines; 19 U.S. targets spanned pharmaceutical, semiconductor, defense, nuclear, and financial sectors.
- CrowdStrike told CyberScoop that ~70% of targets were confirmed compromised or had moderate-confidence evidence of intrusion during the observed period.
- Recovered transcripts showed a human-in-the-loop operator issuing instructions in Chinese (including tasks like writing web shells, extracting domain admin passwords, and checking for hosts lacking the CrowdStrike Falcon sensor).
- CrowdStrike reported Vapor Panda deployed the BLOODALCHEMY backdoor in at least some intrusions and published C2 indicators including nginxmysql[.]com, sqlserver-sqlite[.]com, and www[.]ddduuuppp[.]com (plus associated IPs in the report).
- CrowdStrike assessed this activity is separate from Microsoft’s ARTEX-linked Storm-0940 reporting, suggesting multiple China-nexus actors are adopting the framework.
Next Steps
- Block and hunt for BLOODALCHEMY C2: nginxmysql[.]com (45.153.126[.]223), sqlserver-sqlite[.]com (89.46.234[.]50), and www[.]ddduuuppp[.]com (192.121.47[.]20).
- Hunt for ARTEX artifacts called out by CrowdStrike, including command-line parameters containing “artx”/“artex,” “artx-” prefixed cloud roles, “art18-” administration-command comments, and the C:\ProgramData\artx5v\ directory.
- Rotate signing/pre-auth keys and service credentials after any suspected data exposure (including historical backups/directory exports), as described in CrowdStrike’s recommendations.
Read more at Cyberscoop, The Hacker News, Talkback.sh
Zero-click Android “SMS of Death” WAP Push bug can crash telephony via a single malformed data SMS; patch landed quietly in October 2026 updates
Researchers published a proof-of-concept for an Android telephony flaw where a specially crafted WAP Push (data SMS) triggers a crash of the com.android.phone process, dropping calls and disrupting cellular service without any user interaction. Google’s AOSP fix adds exception handling and cleanup during WAP PDU decoding to prevent persistent crash loops, but reports indicate the issue was not clearly listed in the October 2026 Android security bulletin and some older releases may remain unpatched.
Key Details
- Reported testing showed repeated sends (e.g., ~20–30 messages) can escalate impact, with claims of reboot loops and persistent disruption attributed to repeated processing/redelivery and SMS-store issues.
- The bug is in WapPushOverSms.decodeWapPdu(), where a header-length value is converted to a signed 32-bit integer without proper bounds checking, leading to a NegativeArraySizeException when allocating an array.
- Because it arrives as a data SMS (WAP Push) processed before the inbox, the message may leave no visible SMS entry or notification for the recipient, per the reports.
- The issue is tracked in reporting as CVE-2026-58882, with the patch referenced as included in October 2026 updates for Android 14–17 in the coverage.
Next Steps
- Verify fleet patch levels for the October 2026 Android security update (or OEM equivalent), prioritizing devices that handle corporate voice/SMS workflows and any Android 12/13 devices cited as potentially not receiving the fix.
Read more at android.googlesource.com, Bleeping Computer, Cyber Security News, Talkback.sh
Leaked SquareML access key left 238,083 Denver Health Medical Plan members’ data potentially exposed for over a year
Denver Health Medical Plan (DHMP) disclosed that a SquareML vendor access key was found on a public website and may have been used to access a SquareML database containing data on 238,083 DHMP members. The key was tied to a read-only account and DHMP says it has no evidence of viewing or misuse, but the database contained member personal details and health-related information.
Key Details
- SquareML said the key was inadvertently published on June 27, 2025, and it was deactivated on Aug. 18, 2026 after DHMP was notified on Aug. 14, 2026.
- Data present in the SquareML-managed database included names, dates of birth, city/state/ZIP, and demographic information.
- The dataset also included health metrics (BMI, HbA1c, eGFR) and plan/provider and hospital admission/discharge/transfer data.
- DHMP reported the database did not include Social Security numbers or financial information, and it reset passwords for all accounts with database access and set up a hotline (888-766-0023) for affected individuals.
- DHMP says: at this time, DHMP has no evidence that member information was viewed or misused, only that it may have been.
Next Steps
- If you have employees that might be using DHMP warn them about potential upcoming targeted phishing attempts related to the DHMP services and data.
- Make sure your servers and services have adequate access logging to be able to detect if and when data was accessed in case of a credential leak.
Read more at denverhealthmedicalplan.org, Security Week
Apple warns users in 99 countries of mercenary spyware targeting ahead of Pall Mall Process standards talks
Apple sent a new round of threat notifications to targeted users in 99 countries, warning of mercenary spyware attacks affecting people including journalists and documentary filmmakers. The company said it does not attribute these alerts to specific attackers or regions, and researchers note the notifications often trigger follow-on investigations by groups like Access Now and Citizen Lab.
Key Details
- Apple said it has now notified users in over 150 countries in total about mercenary spyware targeting.
- Apple described mercenary spyware as extremely costly, sophisticated, and global, and said it is among the most advanced digital threats it tracks.
- Citizen Lab’s John Scott-Railton cautioned that not receiving an Apple notification doesn’t rule out infection, and warned users to be wary of companies selling paid “fixes.”
- Apple’s recommended mitigations highlighted by researchers include enabling Lockdown Mode and contacting the digital rights group Access Now for help.
- The notifications land shortly before final negotiations on voluntary Pall Mall Process industry standards for commercial hacking tools, following an earlier set of guidelines aimed at governments.
Next Steps
- If you receive an Apple threat notification, enable Lockdown Mode on affected devices and follow Apple’s guidance.
- **Consider enabling Lockdown Mode for high risk executives, **as it shrinks a device’s attack surface by restricting attachments, link previews, and other common spyware entry points.
Read more at Cyberscoop
Subscribe
Subscribe to receive this weekly cybersecurity news summary to your inbox every Monday.